Tissue Response to Interventional Mri-Guided

TISSUE RESPONSE TO INTERVENTIONAL MRI-GUIDED

THERMAL ABLATION THERAPY

MICHAEL SCOTT BREEN

Submitted in partial fulfillment of the requirements

For the degree of Doctor of Philosophy

Dissertation Advisor: Dr. David L. Wilson

Department of Biomedical Engineering

CASE WESTERN RESERVE UNIVERSITY

May, 2004

CASE WESTERN RESERVE UNIVERSITY

SCHOOL OF GRADUATE STUDIES

We hereby approve the dissertation of

______

candidate for the Ph.D. degree *.

(signed)______(chair of the committee)

______

(date) ______

*We also certify that written approval has been obtained for any proprietary material contained therein.

To my wife, Miyuki, who has blessed my life with her love, patience, and tremendous understanding

ii TABLE OF CONTENTS

List of Tables………………………………………………………………………….…iv

List of Figures……………………………………………………………………….....…v

Acknowledgements………………………………………………………………….….vii

List of Abbreviations…………………………………………………………………..viii

Abstract…………………………………………………………………………….……..x

Chapter 1: Introduction, Background, and Significance...... 1

Chapter 2: Three Dimensional Method for Comparing MR Images of Thermally

Ablated Tissue with Tissue Response...... 13

Chapter 3: Radio Frequency Thermal Ablation: Correlation of Hyperacute

MR Lesion Images with Tissue Response...... 42

Chapter 4: Registration of Subacute MR Lesion Images to Histological Sections

with Model-Based Evaluation...... 70

Chapter 5: Laser Thermal Ablation: Model and Parameter Estimates to Predict

Cell Death from MR Thermometry Images ...... 101

Chapter 6: Conclusions and Future Work ...... 131

Appendix 1: Principles in Magnetic Resonance Imaging...... 141

Bibliography ...... 148

iii LIST OF TABLES

3.1 Comparison of interobserver variability to MR-histology boundary differences.....60

5.1 Laser heating duration, maximum lesion temperature, and lesion area...... 115

5.2 Estimated parameters...... 117

iv LIST OF FIGURES

1.1 Interventional MRI suite...... 3

2.1 Hyperacute post-ablation MR lesion images ...... 19

2.2 Tissue slicing apparatus and process...... 23

2.3 Histology lesion images...... 26

2.4 Registration quality of MR, tissue, and histology images...... 30

2.5 Plot of histology-tissue registration error...... 31

2.6 Plot of MR-tissue registration error ...... 31

2.7 Plot of MR-histology registration error...... 32

2.8 Zonal response of thermal lesion...... 34

2.9 Comparison of tissue damage boundary with MR image features...... 35

3.1 Hyperacute post-ablation MR lesion images...... 53

3.2 Hyperacute zonal response in histology images...... 55

3.3 Comparison of tissue damage boundaries with MR hyperintense rim...... 56

3.4 Plot of signed distances between histology and MR lesion boundaries...... 58

3.5 Plot of absolute and signed distances between histology and MR boundaries...... 59

4.1 Registration quality for MR, tissue, and histology images...... 86

4.2 Comparison of histology and MR boundaries...... 87

4.3 Fit of ellipsoid model to segmented lesion boundaries...... 88

4.4 Plot of absolute distances from manually segmented boundaries to model...... 89

4.5 Fusion of MR and histology ellipsoidal model surfaces...... 90

4.6 Plot of signed and absolute distances between model surfaces...... 91

4.7 Correspondence of cell death boundary and MR hyperintense rim...... 92

v 4.8 Plot of signed and absolute distances between model surfaces...... 93

5.1 MR temperature maps...... 116

5.2 Registration quality of gradient echo and post-ablation MR lesion images...... 117

5.3 Comparison of cell death boundary with modeled tissue damage map...... 118

5.4 Plot of areas under ROC curves...... 119

5.5 Difference map...... 120

5.6 Plot of number of false positives and negatives for each lesion...... 121

5.7 Plot of number of false positives and negatives for each weight...... 121

5.8 Plot of specificity and sensitivity...... 122

5.9 Plot of distance between cell death boundary and mislabeled voxels ...... 123

5.10 Plot of comparison between our model and critical temperature model...... 123

vi ACKNOWLEDGEMENTS

I would like to thank my advisor, Dr. David L. Wilson, for his guidance and mentorship.

His expertise in medical imaging, emphasis on rigorous experimental methodology and

data analysis, and helpful tutoring for scientific paper presentations provided me with invaluable research experience.

I would like to thank my committee members, Drs. Jonathan S. Lewin, Gerald M.

Saidel, Maryann Fitzmaurice, Jeffrey L. Duerk, and Frank L. Merat for their guidance and many helpful suggestions. I would like to acknowledge the contributions of Dr. Roee

S. Lazebnik, a previous student in Dr. Wilson’s lab, who provided his expertise in MR image-guided ablation studies and is coauthor on journal papers represented by Chapters

2, 3, and 4 of this dissertation. I also acknowledge the financial support of the NIH through grant number R01-CA84433 and training award number GM07535-25.

I would like to deeply thank my wife, Miyuki, for her expertise in statistics, helpful advice in mathematical modeling, and continual support and encouragement.

vii LIST OF ABBREVIATIONS

2D: Two Dimensional

3D: Three Dimensional

CE: Contrast Enhanced

CT: X-ray Computed Tomography

FN: False Negative

FOV: Field Of View

FP: False Positive

FPF: False Positive Fraction

GE: Gradient Echo

GRE: Gradient Recalled Echo

GUI: Graphical User Interface

H&E: Hematoxylin and Eosin

HIFU: High Intensity Focus Ultrasound

iMRI: Interventional Magnetic Resonance Imaging

LITT: Laser Induced Interstitial Thermal Therapy

MR: Magnetic Resonance

MRI: Magnetic Resonance Imaging

MT: Masson Trichrome

NSA: Number of Acquisitions

PET: Positron Emission Tomography

PRF: Proton Resonance Frequency

viii RF: Radio Frequency

ROC: Receiver Operating Characteristics

ROI: Region of Interest

SE: Spin Echo

SNR: Signal to Noise Ratio

SPECT: Single Positron Emission Computed Tomography

T1: spin-lattice relaxation time of tissue

T2: spin-spin relaxation time of tissue

TE: Echo Time

TN: True Negative

TP: True Positive

TPF: True Positive Fraction

TPS: Thin Plate Spline

TR: Repetition Time

ix Tissue Response to Interventional MRI-guided Thermal Ablation Therapy

Abstract

MICHAEL SCOTT BREEN

This research project is part of a larger, long-term effort to develop a minimally invasive and cost effective method to ablate solid tumors using interventional MRI (iMRI) to guide and monitor therapy. A low-field, open magnet system is used to guide an ablation probe into the tumor and to monitor tissue destruction during the ablation procedure.

Ablation may be produced by heating tissue from a radio frequency (RF) or laser energy source at the probe tip. Not only does MR provide tumor visualization, it can reveal the thermal lesion in various acquisition sequences (T2-weighted and T1-weighted with gadolinium contrast agent) and measure temperature changes. Potentially, these measurements can be used during clinical ablation procedures to determine tumor cell death and to minimize damage to normal surrounding tissue of critical importance. This research addresses how well these MR measurements predict the actual tissue response.

In this work, we addressed this issue using animal models. We developed experimental and computer techniques to accurately register and correlate MR images with macroscopic tissue and histology images showing tissue response. To ascertain cell death or injury, we used various histological techniques. Ultimately, our goal was to

x quantitatively predict cell damage and death using MR image acquisition and analysis

methods, and a cell death model that accounts for the tissue response from the

temperature history. With this research, we established an analysis paradigm suitable for many future studies of ablation techniques, localized drug release, and other iMRI-guided therapies.

xi Chapter 1

Introduction, Background, and Significance

1.1 Introduction

Interventional image-guided ablation therapies using thermal energy sources such as radio-frequency (RF), microwave, laser, high-intensity focused ultrasound (HIFU), and cryogenics have received much recent attention as minimally invasive strategies for the treatment of localized malignant disease (1-3). Potential benefits of these techniques include the ability for near real-time image guidance using magnetic resonance (MR),

X-ray computed tomography (CT), or ultrasound, the ability to ablate tumor in nonsurgical candidates, and the potential to perform the procedure on an outpatient basis.

Further, it is believed that image-guided ablation therapy may eventually compete with some open tumor surgeries (4-6).

1.2 Magnetic Resonance Imaging

MR imaging (MRI) is a rapidly growing and changing image modality. The benefits of

MRI include the use of nonionizing radiation, excellent soft tissue discrimination, sensitivity to blood flow and temperature, and ability to image at any angle. Significant improvements in image quality with increasing complex pulse sequences and improved equipment designs have been key reasons for the expanding clinical use of MRI. There are drawbacks, including expensive equipment, relatively long imaging times, and patient

1 claustrophobia issues. The basic principles of MR imaging and thermometry are described in the Appendix 1.

1.3 MR Image-Guided Thermal Ablation Therapy

RF tumor ablation has been used in early clinical trials for the treatment of liver, kidney, prostate, pancreas, and musculoskeletal system tumors (7). In a conventional interventional MRI (iMRI)-guided RF ablation performed with a system as shown in

Figure 1.1, a small diameter needle electrode with a 10-20 mm exposed tip is inserted percutaneously into the target tissue using MR for near real-time guidance. Lesion formation is achieved by increasing the local tissue temperature via resistive heating with the delivery of RF electric current between the electrode tip and ground pads which are placed on the patient’s skin. The ablation is simply controlled by adjusting the power of a

500 kHz RF generator. The tip of the RF electrode is typically maintained at a constant temperature using a thermistor within the electrode tip to provide accurate temperature information. The resistive tissue heating induces cellular death via coagulative necrosis or other mechanisms (8). The lesion volume is determined by several factors including the duration of heating, length of electrode’s exposed tip, and perfusion mediated tissue cooling.

Laser-induced interstitial thermal therapy (LITT), has been applied for the minimally invasive treatment of solid tumors (9-12). LITT requires only the percutaneous insertion of an optical fiber with a light diffuser tip through a needle into the pathological tissue under image guidance. Laser energy, from a neodymium:yttrium aluminum garnet

(Nd:YAG) source or others, delivered through the fiber is absorbed by tissue proximal to

2 the diffuser tip, resulting in a local temperature increase and thermal destruction.

A C

Figure 1.1. An interventional MRI suite consisting of an open low-field (0.2T) MR imaging system (A), in-room MR-compatible monitor with mouse to control the system (B), and RF thermal lesion generator (C). The open bore allows access to patient for interactive image-guided placement of the RF electrode into the pathological tissue. RF current is applied to ablate the tissue by resistive heating.

1.4 MR and Tissue Responses to Thermal Ablation

There have been some inconsistent findings with regard to tissue damage at the margin of

MR images of thermal lesions acquired minutes after treatment. Some studies suggested that a central region of low MR signal corresponds to the region of tissue damage (13-

16). For example, in normal rabbit liver, Lee et al. (13) manually measured diameters of the central hypointense region in T2-weighted and gadolinium contrast-enhanced (CE)

T1-weighted MR thermal lesions, and measured diameters of a central region of color change in sectioned fresh tissue. In six lesions, the diameters agreed within 2 mm. Other studies of ablation showed that a hyperintense rim in MR may become necrotic

3 (17;18). Merkle et al. (17) compared the diameters of a region occupied by the union of the central zone and surrounding hyperintense zone in T2 and CE T1-weighted MR lesion images with diameters of coagulation seen in fixed macroscopic tissue slices, without any correction for tissue shrinkage. These diameters matched within typically

2 mm. From these studies, there is evidence that MR thermal lesion images reflect tissue damage. However, the inconsistencies in the literature indicate the challenge of identifying cell death from color changes in macroscopic tissue sections and from morphological changes in hematoxylin and eosin (H&E) stained histology immediately following ablation. Another potential source of discrepancy is the limitations of diameter measurements for potentially nonsymmetrical lesion boundaries. A careful regional correlation requires the alignment of histology to MR images, and reliable methods to accurately determine the extent of tissue damage. Our studies carefully examined the relationship between cell damage and MR images on a voxel-by-voxel basis.

There are reports of using MR temperature measurements to monitor ablation (19-

21). Chen et al. (19) compared MR temperatures to MR lesion images of in vivo rabbit brain following laser ablation. The border of tissue damage corresponded to a threshold temperature of approximately 48°C. Peters et. al. (21) determined that a threshold at 51°C or a thermal dose of 200 minutes in the in vivo canine prostate could predict the thermal-injury border in post-ablation MR lesion images. The model for these studies assumed an Arrhenius function above 43°C for the thermal dose calculation, and utilized only temperature data at the lesion margin. However, the empirically-derived values for these Arrhenius-based models can vary for different tissues, temperature ranges, and heating durations. In this work, we developed a new tissue damage model and applied it

4 on a voxel-by-voxel basis to temperature-time histories as obtained from multiple MR thermometry measurements.

1.5 Image Registration of MR Images to Tissue Response

There are various potential experimental approaches for spatially correlating MR and tissue response. Previous reports have described various two dimensional (2D) methods for correlating MR and tissue response (18;22). Morrison et al. (22) used a 2D method that required the insertion of a second needle parallel to the ablation needle, and image acquisition and tissue slicing in the plane of both needles. We find this difficult for some organs such as an in vivo kidney. Also, it is difficult to exactly orient an organ and slice it in the correct plane. Further, when one obtains slices parallel to the needle, tissue in histology opens at the needle track and losses its geometric integrity. An alternative 2D method is to slice the tissue perpendicular to the needle. Orienting the tissue slice exactly perpendicular to the needle is again difficult. Another 2D method used by Chen et al. (18) created two lesions in a plane parallel to the sagittal plane containing the longitudinal fissure (separation between the right and left cerebral hemispheres), and the distance between these planes was measured in MR images. By slicing parallel to the longitudinal fissure and at the measured distance, the plane containing the middle of the lesions was sectioned. For these 2D approaches, the experiment is lost with no possible recovery when the plane of interest is missed in cutting. We addressed these limitations in our research by developing a three dimensional (3D) technique for relating histology to MR images.

1.6 Significance

This research is significant because the successful application of interventional technology to minimally invasive treatment of localized pathologies has the potential to prolong life while reducing the morbidity and cost associated with a more invasive surgical approach (2;3).

As compared with other imaging modalities, there are several advantages to

MR-guided thermal ablation. MR imaging has a lack of ionizing radiation, excellent soft tissue contrast, high spatial resolution, directly acquired multiplanar scan plane, and sensitivity to blood flow and temperature (23-25). This not only permits accurate destruction of the tumor, it extends the application of ablation to the safe destruction of tumor adjacent to vital structures. In addition, MR imaging appears to be more accurate than ultrasound or CT for monitoring the effects of thermal ablation treatment (7;26). The major contribution of MRI is its potential to monitor the zone of thermal tissue destruction during the procedure. To monitor an ablation procedure, MRI can intermittently acquire temperature images during heating and structural lesion images during and after heating. By comparing MR temperature measurements and thermal lesion images to tissue response, this research investigated this most important aspect of

MRI-guided ablation.

The research and development work in this research addresses one of the most fundamental questions in iMRI ablation: “Can MR measurements predict cell death in the treated region with minimal damage to surrounding normal tissue?”. Although changes in

MR lesion images and temperature measurements are observed and used to adjust

6 therapy, the actual tissue response is required. In this work, we demonstrate that the MR signal accurately reflects tissue damage and destruction. This will enable one to adjust the size and shape of the treated region, through additional heating or displacement of the RF electrode to a new location, so as to ensure coverage of the pathology. Alternatively, in the case of tumor adjacent to vital structures such as the gall bladder, bowel, and especially the brain, the physician can monitor zones of cell damage during the ablation.

A display of color-coded zones of cell death and damage predicted from MR thermometry images and a tissue damage model will slowly develop allowing the physician to stop well before critical structures are damaged.

1.7 Organization

The main goal of this research is to quantitatively predict tissue damage and death using

MR image acquisition and analysis methods, and a model that accounts for the tissue response to temperature history. To achieve this goal, we developed three dimensional methodologies to accurately map tissue response to in vivo MR images, evaluated cell death using MR thermal lesion images, and established a mathematical model to predict tissue damage from a sequence of MR temperature images.

Each chapter in this dissertation (Chapters 2 through 5) consists of a manuscript that has been submitted or published in a peer-reviewed journal. Chapter 2 introduces a three dimensional registration method to align in vivo MR thermal lesion images with macroscopic tissue and histology images. Chapter 3 investigates the hypothesis that the outer boundary of the hyperintense region observed in hyperacute (several minutes post-ablation) T2 and contrast-enhanced T1-weighted MR lesion images is an accurate

7 predictor of eventual cell death from RF thermal ablation. Chapter 4 further examines this relationship with studies performed on animals sacrificed four days post-ablation to unambiguously reveal the complete necrotic region in histology. Chapter 5 describes our development and evaluation of a model and parameter estimates to predict cell death from a sequence of MR thermometry measurements.

In Chapter 2, we introduce a new methodology using three-dimensional registration to spatially correlate in vivo MR and histology images. Our method requires both computer registration methods and special tissue handling techniques. We present several validation results showing the accuracy of the method, and describe a preliminary comparison of MR and tissue response.

In Chapter 3, we test the hypothesis that the outer border of the hyperintense region observed in hyperacute T2 and CE T1-weighted MR lesion images following RF ablation is an accurate predictor of eventual cell death. To investigate this idea, we used several specially designed experimental and registration techniques so as to enable alignment of MR lesion images with histological images showing tissue damage and destruction.

To unequivocally reveal the necrotic region in histology, we performed experiments with animals sacrificed four days post-ablation. In Chapter 4, we used an ellipsoidal model, which describes the thermal lesion surfaces, to correlate the cell death region in histology with T2-weighted MR images four days post-ablation.

In Chapter 5, we developed a tissue damage model to predict cell death from MR thermometry measurements, and applied it on a voxel-by-voxel basis to in vivo rabbit brain data. Our method included image registration, temporal filtering of MR temperature

8 image sequence, and parameter estimation techniques. Model parameters were simultaneously estimated using an iterative optimization algorithm applied to every interesting voxel in hundreds of images from multiple experiments having various temperature histories.

Finally, in Chapter 6, we summarize the main conclusions and the original contributions of this research. We also propose future work to be investigated.

9 WORKS CITED

1. Goldberg, S. N.; Livraghi, T.; Solbiati, L.; Gazelle, G. S. In situ ablation of focal hepatic neoplasms. Gazelle, G. S., Saini S., and Mueller, P. R. Hepatobiliary and pancreatic radiology: imaging and intervention. New York: Thiema Medical Pub; 1997.

2. Jolesz, F. A. and Shtern, F. The Operating Room of the Future. Report of the National Cancer Institute Workshop, "Imaging-Guided Stereotactic Tumor Diagnosis and Treatment". Invest Radiol 1992;27(4):326-8.

3. Jolesz, F. A. and Blumenfeld, S. M. Interventional Use of Magnetic Resonance Imaging. Magn Reson.Q. 1994;10(2):85-96.

4. Jolesz, F. A., Bleier, A. R., Jakab, P., Ruenzel, P. W., Huttl, K., and Jako, G. J. MR Imaging of Laser-Tissue Interactions. Radiology 1988;168(1):249-53.

5. Kahn, T., Bettag, M., Ulrich, F., Schwarzmaier, H. J., Schober, R., Furst, G., and Modder, U. MRI-Guided Laser-Induced Interstitial Thermotherapy of Cerebral Neoplasms. J Comput.Assist.Tomogr. 1994;18(4):519-32.

6. Vogl, T. J., Mack, M. G., Muller, P., Phillip, C., Bottcher, H., Roggan, A., Juergens, M., Deimling, M., Knobber, D., Wust, P., and . Recurrent Nasopharyngeal Tumors: Preliminary Clinical Results With Interventional MR Imaging--Controlled Laser- Induced Thermotherapy. Radiology 1995;196(3):725-33.

7. Gazelle, G. S., Goldberg, S. N., Solbiati, L., and Livraghi, T. Tumor Ablation With Radio-Frequency Energy. Radiology 2000;217(3):633-46.

8. Cosman, E. R., Nashold, B. S., and Ovelman-Levitt, J. Theoretical Aspects of Radiofrequency Lesions in the Dorsal Root Entry Zone. Neurosurgery 1984;15(6):945-50.

9. Mack, M. G. and Vogl, T. J. MR-Guided Ablation of Head and Neck Tumors. Magn Reson.Imaging Clin.N.Am. 2002;10(4):707-13, vi.

10. Izzo, F. Other Thermal Ablation Techniques: Microwave and Interstitial Laser Ablation of Liver Tumors. Ann.Surg.Oncol. 2003;10(5):491-7.

11. Muralidharan, V., Malcontenti-Wilson, C., and Christophi, C. Interstitial Laser Hyperthermia for Colorectal Liver Metastases: the Effect of Thermal Sensitization and the Use of a Cylindrical Diffuser Tip on Tumor Necrosis. J Clin.Laser Med.Surg. 2002;20(4):189-96.

12. Hall-Craggs, M. A. and Vaidya, J. S. Minimally Invasive Therapy for the Treatment of Breast Tumours. Eur.J Radiol 2002;42(1):52-7.

10 13. Lee, J. D., Lee, J. M., Kim, S. W., Kim, C. S., and Mun, W. S. MR Imaging- Histopathologic Correlation of Radiofrequency Thermal Ablation Lesion in a Rabbit Liver Model: Observation During Acute and Chronic Stages. Korean J.Radiol. 2001;2(3):151-8.

14. Merkle, E. M., Shonk, J. R., Zheng, L., Duerk, J. L., and Lewin, J. S. MR Imaging- Guided Radiofrequency Thermal Ablation in the Porcine Brain at 0.2 T. Eur.Radiol. 2001;11(5):884-92.

15. Merkle, E. M., Haaga, J. R., Duerk, J. L., Jacobs, G. H., Brambs, H. J., and Lewin, J. S. MR Imaging-Guided Radio-Frequency Thermal Ablation in the Pancreas in a Porcine Model With a Modified Clinical C-Arm System. Radiology 1999;213(2):461-7.

16. Boaz, T. L., Lewin, J. S., Chung, Y. C., Duerk, J. L., Clampitt, M. E., and Haaga, J. R. MR Monitoring of MR-Guided Radiofrequency Thermal Ablation of Normal Liver in an Animal Model. J.Magn Reson.Imaging 1998;8(1):64-9.

17. Merkle, E. M., Boll, D. T., Boaz, T., Duerk, J. L., Chung, Y. C., Jacobs, G. H., Varnes, M. E., and Lewin, J. S. MRI-Guided Radiofrequency Thermal Ablation of Implanted VX2 Liver Tumors in a Rabbit Model: Demonstration of Feasibility at 0.2 T. Magn Reson.Med. 1999;42(1):141-9.

18. Chen, L., Bouley, D. M., Harris, B. T., and Butts, K. MRI Study of Immediate Cell Viability in Focused Ultrasound Lesions in the Rabbit Brain. J.Magn Reson.Imaging 2001;13(1):23-30.

19. Chen, L., Wansapura, J. P., Heit, G., and Butts, K. Study of Laser Ablation in the in Vivo Rabbit Brain With MR Thermometry. J Magn Reson.Imaging 2002;16(2):147-52.

20. McDannold, N. J., King, R. L., Jolesz, F. A., and Hynynen, K. H. Usefulness of MR Imaging-Derived Thermometry and Dosimetry in Determining the Threshold for Tissue Damage Induced by Thermal Surgery in Rabbits. Radiology 2000;216(2):517-23.

21. Peters, R. D., Chan, E., Trachtenberg, J., Jothy, S., Kapusta, L., Kucharczyk, W., and Henkelman, R. M. Magnetic Resonance Thermometry for Predicting Thermal Damage: an Application of Interstitial Laser Coagulation in an in Vivo Canine Prostate Model. Magn Reson.Med. 2000;44(6):873-83.

22. Morrison, P. R., Jolesz, F. A., Charous, D., Mulkern, R. V., Hushek, S. G., Margolis, R., and Fried, M. P. MRI of Laser-Induced Interstitial Thermal Injury in an in Vivo Animal Liver Model With Histologic Correlation. J.Magn Reson.Imaging 1998;8(1):57-63.

23. Schenck, J. F., Jolesz, F. A., Roemer, P. B., Cline, H. E., Lorensen, W. E., Kikinis, R., Silverman, S. G., Hardy, C. J., Barber, W. D., Laskaris, E. T., and .

11 Superconducting Open-Configuration MR Imaging System for Image-Guided Therapy. Radiology 1995;195(3):805-14.

24. Cline, H. E., Schenck, J. F., Watkins, R. D., Hynynen, K., and Jolesz, F. A. Magnetic Resonance-Guided Thermal Surgery. Magn Reson.Med. 1993;30(1):98- 106.

25. Cline, H. E., Hynynen, K., Watkins, R. D., Adams, W. J., Schenck, J. F., Ettinger, R. H., Freund, W. R., Vetro, J. P., and Jolesz, F. A. Focused US System for MR Imaging-Guided Tumor Ablation. Radiology 1995;194(3):731-7.

26. Lewin, J. S., Connell, C. F., Duerk, J. L., Chung, Y. C., Clampitt, M. E., Spisak, J., Gazelle, G. S., and Haaga, J. R. Interactive MRI-Guided Radiofrequency Interstitial Thermal Ablation of Abdominal Tumors: Clinical Trial for Evaluation of Safety and Feasibility. J.Magn Reson.Imaging 1998;8(1):40-7.

CHAPTER 2

Three Dimensional Method for Comparing MR Images of Thermally Ablated Tissue with Tissue Response

2.1 INTRODUCTION

Solid tumors and other pathologies are treated using radio-frequency (RF) ablation under

interventional MRI (iMRI) guidance. There are many advantages of iMRI including

excellent soft tissue contrast, high vascular conspicuity, lack of ionizing radiation, and

the ability to image at any angle. To monitor ablation with iMRI, one can measure

temperature during the procedure and/or MR signal amplitudes intermittently during and

following treatment. We are investigating this unique ability to monitor treatment by

comparing MR images of the thermal lesion to cellular damage as seen histologically

using new three-dimensional (3D) registration techniques. If MR measurements can

accurately predict regions of cell death and damage, iMRI will be quite advantageous for

optimally treating tumor while sparing nearby normal tissues of critical importance.

Previous studies indicate that regions of low MR signal in images obtained minutes

after thermal ablation correlate with regions of tissue damage (1-7). Analysis was

typically done using geometric measurements without alignment of histology to MR

images (3;4;6). For example, in normal in vivo rabbit liver, Boaz et al. (4) manually

measured diameters of MR thermal lesions (T2, STIR, and T1 with gadolinium contrast agent, hereafter called Gd) and measured diameters of color changes in sliced fresh tissue. In six lesions, diameters agreed within 2 mm. The typical response was a central

hypointense region surrounded by a hyperintense rim. Lewin et al. (6) manually

estimated axial dimensions and used an ellipsoidal model to estimate tumor and thermal

lesion volumes in humans with no tissue confirmation. From these studies, there is

evidence that MR thermal lesion images reflect tissue damage. However, a careful

regional correlation requires the alignment of histology to MR images. Our studies will

carefully examine the relationship between tissue damage and MR images on a

voxel-by-voxel basis.

There are various potential experimental approaches for spatially correlating MR and

tissue response. Previous reports have described various 2D methods for correlating MR

and tissue response (5;7). Morrison et al. (5) used a 2D method that required the insertion

of a second needle parallel to the ablation needle, and image acquisition and tissue slicing

in the plane of both needles. We find it difficult for some organs such as an in vivo

kidney. Also, it is difficult to exactly orient a kidney and slice it in the correct plane.

Further, when one obtains slices parallel to the needle, tissue in histology opens at the needle track and losses its geometric integrity. An alternative 2D method is to slice the tissue perpendicular to the needle. Orienting the tissue slice exactly perpendicular to the needle is again difficult. Another 2D method used by Chen et al. (7) created two lesions in a plane parallel to the sagital plane containing the longitudinal fissure (separation between the right and left cerebral hemispheres), and the distance between these planes was measured in MR images. By slicing parallel to the longitudinal fissure and at the measured distance, the plane containing the middle of the lesions was sectioned. For these 2D approaches, the experiment is lost with no possible recovery when the plane of

interest is missed in cutting. Our 3D technique for relating histology to MR images addresses these limitations.

A 3D method for aligning histology and MR images is theoretically better and provides more detail than the other methods reviewed above. Using our 3D method, we can address the practical clinical problem of determining if MR images of thermal lesion correlate with tissue response in tissue containing tumor. For separate geometric measurements in MR images and sliced tissue, slicing a specimen in the proper plane is difficult. Also, geometric measurements are limited to normal tissue with symmetrical boundaries. Alignment of histology and MR images is necessary for tumor tissue with irregular boundaries. Additionally, we have observed nonsymmetrical thermal lesion boundaries in normal muscle tissue. For 2D alignment methods, orienting and slicing an organ in the correct plane is problematical, and a technique for correcting misaligned slices is unclear. We used a 3D alignment method to alleviate these problems.

In addition to this iMRI study, the methodology that we are developing should apply to other investigations where one wishes to compare 3D medical images to excised tissue or histology. We are investigating its use in modeling the tissue response from temperature history as obtained from iMRI thermal images; in nuclear medicine, antibody imaging of human prostate cancer (8); and in the characterization of a localized, controlled drug release device (9). As new methods for 3D imaging, including molecular imaging, are developed, we believe that comparisons to histology will be crucial for their characterization and validation.

In the present study, we introduce the method and correlate in vivo MR thermal lesion images to tissue response in an animal model consisting of rabbit thigh muscle.

Our 3D method requires both computer registration methods and tissue handling techniques as described in the next section. We present several validation results showing the accuracy of the method. A preliminary comparison of MR and tissue response is then described.

2.2 MATERIALS AND METHODS

2.2.1 Methods for RF Ablation Experiments

The correlation of in vivo MR and histological image data required careful animal and tissue handling methods. The experimental methods included tissue ablation, MR imaging, tissue slicing and photographing, and histological processing; all were very important for successful registration. It was extremely important to minimize tissue deformation and destruction during the dissection and slicing processes for accurate registration. The details and the validation experiment for the registration are described in the next sections.

2.2.2 RF Ablation in Rabbit Model and MR Imaging

Following a protocol approved by the Institutional Animal Care and Use Committee, we anesthetized 10 New Zealand White rabbits (3.0-3.5 kg) with a 2 ml IM injection of a cocktail (0.6 ml/kg) – a combination of Ketamine (0.26 ml/kg), Xylazine (0.26 ml/kg), and Acepromazine (0.08 ml/kg) (each manufactured by Phoenix Scientific, St. Joseph,

MO). Approximately 30 minutes later, we administered another 2 ml IM injection of cocktail (0.6 ml/kg). Thereafter, IM injections were alternated between 0.5 ml of ketamine (0.15 ml/kg) and 1 ml of cocktail (0.3 ml/kg) every 20-40 minutes for

maintenance of anesthesia. In addition, a 1 ml IM injection of cocktail (0.3 ml/kg) was administered 1 minute before ablation.

We prepared the rabbits for an RF ablation in the right and left thigh muscle. After shaving each animal’s abdomen, and left and right thigh, we placed the rabbits in the prone position within the gantry of a clinical 0.2 T C-arm MR imaging system (Siemens

MAGNETOM OPEN, Erlangen, Germany). Securing the legs of each rabbit to a customized Plexiglas support prevented movement of thighs. After positioning two

8 x 12 cm wire mesh grounding pads (Radionics, Burlington, MA) coated with conductive gel (Aquasonic 100: Parker Laboratories, Orange, NJ) on each rabbit’s abdomen, we placed the thighs within a 12 cm diameter multi-turn solenoid, receive-only coil.

For the thigh muscle facing the front of the MR imaging system, we performed the

RF ablation procedure under MR guidance by using an in-room 1024 x 1280 pixel

RF-shielded liquid crystal monitor and an MR-compatible mouse-driven system (each manufactured by Siemens, Erlangen, Germany). To determine an ablation site (a volume sufficiently distant from major vessels and bone), we produced images of three adjacent slices by applying a FISP sequence with TR/TE/number of acquisitions (NSA) parameters of 17.8/8.1/2 to give 128 x 256 x 3 voxels over a 200 x 200 x 15 mm field of view (FOV) to yield 1.56 x 0.78 x 5.0 mm voxels. Images required 6 seconds of scan time to acquire the three contiguous sections. By placing a water-filled syringe on the skin surface, we localized the target. We inserted an MR-compatible 17-G titanium RF electrode with a 10 mm or 20 mm exposed tip (Radionics, Burlington, MA) percutaneously into the thigh muscle. The FISP imaging sequence previously described

was used to monitor electrode insertion in a near real-time fashion. Once the electrode tip

was positioned, lesion formation was achieved by increasing the local temperature by

delivering RF electric current between the electrode tip and ground pads. We applied RF

energy for 2 minutes using a 100 W RF generator operating at 500 kHz (RFG-3C,

Radionics, Burlington, MA). The tip of the RF electrode was maintained at a temperature of 90 ± 2°C using a thermistor within the electrode tip to provide accurate instantaneous temperature information.

Immediately after ablation, we prepared this thigh muscle for imaging. A

22 G x 1 inch IV catheter (Terumo Medical Corporation, Elkton, MD) was inserted in a dorsal ear vein in each rabbit. At least two MR-compatible 22 G, 10 cm fiducial needles

(E-Z-EM, Westbury, NY) were inserted into the thigh near the thermal lesion with one fiducial approximately parallel and the other fiducial approximately at an angle of 45° to the RF electrode. The RF electrode was removed from the thigh before imaging to prevent MR image artifacts.

Approximately 10 minutes after ablation, we acquired MRI volumes of the thigh. We used two different MR sequences. A T2-weighted turbo-spin-echo sequence was applied with TR/TE/NSA parameters of 3362/68/8 gives 256 x 256 x 9 voxels over a

180 x 180 x 27 mm FOV to yield 0.70 x 0.70 x 3.0 mm voxels oriented to give the highest resolution for slices perpendicular to the fiducial needle placed parallel to the RF electrode. This imaging plane minimizes partial volume error because temperature changes little along the needle. We also applied a T1-weighted spin-echo sequence with

TR/TE/NSA parameters of 624/26/6 gives 256 x 256 x 9 voxels over a

180 x 180 x 27 mm FOV to yield 0.70 x 0.70 x 3.0 mm voxels oriented the same as the

T2-weighted sequence. At least 5 minutes before acquiring the T1-weighted spin-echo images, we administered an IV injection of gadolinium contrast medium (0.2 mmol/kg, gadopentate dimeglumine: Berlex Laboratories, Wayne, NJ). Figure 2.1 shows a typical

T2-weighted and gadolinium contrast-enhanced T1-weighted MR image following ablation. After imaging, the rabbits were turned for access to the other thigh, and the same ablation and imaging procedure was repeated.

Figure 2.1. A typical T2-weighted (a) and T1-weighted with gadolinium contrast enhancement (b) MR image acquired within minutes after ablation. The thermal lesion (horizontal arrow) is the bright elliptical region that has a hypointense inner region surrounded by a hyperintense outer region in both images. The small dot at the center of the lesion is the track of the RF electrode, which was withdrawn prior to imaging. The two small dark regions (vertical arrows) are the MR image artifacts of the fiducial needles. In a magnified view of the right needle artifact (inset), a characteristic artifact is shown. There are three small bright spots visible at the edge of the dark region. An overlay is shown of a triangle formed by these three bright spots. The approximate location of the needle’s center (cross) is slightly offset from the center of the triangle towards the pointed end.

We then prepared both thighs for examining the tissue response. To examine latent tissue changes in response to the interventional procedure, the rabbits were sacrificed

45 minutes after the ablation using a barbiturate overdose technique via intravenous administration of 0.1 ml/lb of pentobarbital sodium (Euthasol, 390mg/ml, Diamond

Animal Health Inc., Des Moines, IA). To prevent significant thigh muscle deformation, the entire pelvis and back limbs of the rabbits were harvested and fixed in 10 percent formalin for 2-3 days. The formalin was maintained at approximately 6°C in a refrigerator to prevent any deterioration of the tissue before fixation. To improve the penetration of the formalin, we then removed the thighs from the bone and fixed them for an additional 10-12 days. After the tissue was fixed, we acquired histological samples and registered them to the MR images using the method described in subsequent sections.

2.2.3 Methods for Registering MRI and Histology

To correlate histological and 3D medical image data, we use semi-automatic registration algorithms as well as special tissue handling procedures. Here we briefly review the entire approach so as to introduce the details of the computer registration methods.

Experimental details were described in previous sections.

MR images of the rabbit thigh are obtained. Fiducial needles are included that can be seen in the MR images and that leave a hole that can be found in the tissue. The fiducial needle tracks are used for 3D registration. After the tissue is fixed, we slice it on a special custom-made apparatus into 3 mm slices. Photographs of the tissue block face become our reference tissue images. Histology samples are obtained in a plane parallel to these tissue images and photographed on a video microscopy system. We register both the MR

images and the histology images to the reference tissue images in order to compare MR

and histological measurements.

2.3.4 3D Registration of MRI to Reference Tissue Images

We used a 3D registration algorithm to align fiducial needle paths, and optional point

landmarks, in MR images to those in calibrated tissue images. Briefly, for every slice in

each of the two volumes, we manually selected the center of the needle paths, and

optional point landmarks, to create a list of 3D spatial locations in millimeters. An

automated optimization was then used to determine the best transformation consisting of

rotation, translation and uniform scaling. The iterative closest point algorithm as

described by Besl and McKay (10) was used. The optimal transformation matrix was

used to reslice the MRI volume using trilinear interpolation. We used at least two needles for alignment. The method was previously validated by Lazebnik et al. (11) using computer phantoms and brain tissue images, and errors were estimated to be less than

0.9 mm.

2.2.5 2D Registration of Histology to Reference Tissue Images

We often cut each thick-sectioned tissue slice to fit a 5.0 x 5.0 cm histology-mounting block. Images of these “tissue sub-sections” are obtained using a digital camera and copy stand, and aligned to their corresponding reference tissue images using a 2D affine transformation. In each image, we use a cursor to select at least three correspondence

points such as the RF probe track, fiducial needle tracks, and corner points. We calculate

the least-squares solution and transform the tissue sub-section image to match the tissue

image.

To align the histology image with this transformed tissue sub-section image, we

developed a 2D thin plate spline (TPS) warp algorithm that is described in detail

elsewhere (12). Using a cursor, we choose correspondence points by selecting the interior

ink marks and/or anatomical landmarks such as blood vessels. We also obtain

corresponding points along external boundary segments. Corresponding boundary

segments are obtained by selecting readily identifiable end points consisting of ink marks

or corners. A computer algorithm obtains a curve representing the border segment,

measures the distance between the end points, and creates 5-10 points equidistant along

the curve in both images. These provide additional correspondence points to match. Often

multiple border segments are used. All correspondence points are included to compute

the TPS transformation. The method was validated by Lancaster and Wilson (12).

2.2.6 Tissue Slicing and Acquisition of Calibrated Tissue Images

We examined the tissue response in tissue images and histology slides. As shown in

Figure 2.2, a tissue slicing apparatus was constructed that included a tissue platform and a

digital camera (DSC-D770, Sony, Japan). To prevent any foreshortening in the tissue

photographs, we created a method to ensure that the imaging plane of the camera was

parallel to the tissue face. A straight pipe was placed on the tissue platform perpendicular to the tissue-slicing plane. Using the camera’s display to view the shaft of the pipe, the camera was rotated until the front and back openings of the pipe were centered, and then it was secured with a locking screw.

We developed a technique to slice the tissue in planes approximately perpendicular to

a fiducial needle. By orientating both planes perpendicular to a fiducial needle, we

ensured the tissue-slicing plane and to the plane of the MR image slices were

approximately parallel. This minimizes image blurring of the resliced images when there

is out-of-plane tilting with respect to the thick MR image slices. To orient the specimen

and reduce deformation during slicing, the specimen was positioned in a Styrofoam block

and held in place with tissue embedding wax (50-54°C Parablast X-tra, Oxford Labware,

St. Louis, MO). A black coloring agent was added to the translucent wax to create a clear

tissue border in the photographs. The Styrofoam block was secured to the tissue-slicing

platform with glue and tape.

We obtained 3 mm tissue slices by using the specially designed apparatus that

included a linear displacement device (Rack and Pinion Slide, Edmund Scientific,

Barrington, NJ) for accurate stepping of the platform in small increments. We sliced the

specimen with a 12.8 inch autopsy knife (Tissue-Tek Accu-Edge Semi-Disposable

AutopsyFigure 2.2. Knife Tissue System slicing, Sakura apparatus Finetek, and Japan) process. using The the apparatu verticals supports(a) contains of the a slicingdigital camera, a tissue platform (up arrow), and a linear displacement device (down arrow) for apparatusadvancing asthe a stageguide. in R ae precpeatedly,ise stepped we photogr fashion.aphed The the tissu tissuee sam blockple is face,embedded advanced in wax, the secured to the platform, advanced using the linear displacement device, and sliced using platformthe vertical by uprights3 mm, and as asliced guide th (b).e tissue, The tissueuntil theblock specim facee nis wasthen traversed. photographed The tipswith of a digital camera. This process is repeated until the entire sample is sliced. In a typical tissue fiducialphotograph needles of a rabbitwere thighexposed mu scleat the (c), tissue a rule r,block the tipsface, of highlightedthree fiducial with needles ink for(ver easytical arrows), and the thermal lesion about 1.5 cm in diameter (horizontal arrow) are clearly seen.identificatio A photographn in each of ph tissueotograph, sub-section and stepped (d) shows back thesligh headstly b eofyond ten pinsthe p usedlane ofto theproduce next ink fiducials for registering with histology. The center of the pinheads is identified for tissueexcellent slice. localization of the ink fiducials.

We calibrated these macroscopic tissue images using a ruler in the plane of the tissue slice. To determine the number of pixels per centimeter (pixels/cm) in the images, we used a cursor to select six points on the ruler separated by a distance of 1 cm, and calculated the average Euclidean distance in pixels. This calibration value was stored in a data file. To maintain calibration, the imaging geometry and the manual zoom on the digital camera were fixed during acquisitions.

(Davidson Marking System, Bradley Products, Bloomington, MN), we photographed them on a copy stand. Figure 2.2 shows a typical tissue image containing a lesion and ink fiducials. We embedded the tissue in paraffin and obtained histological samples.

2.2.7 Acquisition of Histological Images

We digitized the histology slides using a video microscopy system with a motorized stage. The system consisted of a light microscope (BX60, Olympus, Japan), video camera

(DXC-390, Sony, Japan), and coupler (U-TV0.35XC, Olympus, Japan), position encoded motorized stage (ProScan, Prior Scientific, Rockland, MA), and software (Image-Pro with Scope-Pro, Media Cybernetics, Silver Spring, MD). To obtain an image of the entire slide, we used the tiling function of Scope-Pro. This procedure drove the motorized stage, acquired a series of pictures, and seamlessly combined the photographs to form one large tiled image. We performed tiling with a 4x objective. Typically, the tiled image consisted of about 100 image acquisitions over a 1.0-1.5 cm lesion and gave about 6400 x 4800

square pixels that were 5.21 um on a side. By reducing the tiled image to 10 percent of its original size, we created a smaller map image on which we marked the locations of ink fiducials and boundaries of the tissue damage. This map image was registered to the tissue image.

Using the video microscopy system and a software-scripting program written in

Image-Pro and Scope-Pro, we determined the location of each needle and ink fiducial on the slide, and marked the appropriate position on the map image. A live video window was displayed with a digitally imposed crosshair at the center of the window. A joystick was used to drive the motors that moved the slide to locate the ink fiducials. While locating fiducials, the operator could switch microscope objectives to acquire images at higher or lower magnification levels. Once a fiducial was located, the operator centered the crosshair over the fiducial and clicked a graphical user interface (GUI) button to obtain the stage coordinates. The fiducial was then marked at the appropriate location on the map image using a colored graphical overlay. Figure 2.3 shows a typical map image containing overlays of ink and needle fiducials, and a video snapshot of a fiducial needle track and an ink fiducial.

We determined the location of each tissue boundary of interest on the slide and marked the appropriate position on the map image using the same video microscopy system and software described previously. With a digitally imposed crosshair at the center of the live video window, we panned around the slide under joystick control and identified boundaries of interest. On each boundary, we identified 15-30 points by centering the crosshair over each point of interest and click a GUI button to acquire the stage coordinates. Each boundary point was marked at the appropriate location on the

map image with a colored graphical overlay. The end result of this marking process was a

map image with tissue boundaries of interest.

Figure 2.3. Typical Masson trichrome stained histology image of rabbit thigh muscle following thermal ablation. A software tiling function was used to move the microscope’s motorized stage, acquire a series of pictures, and seamlessly combine the photographs to form one large image of the entire histology slide (a). The smooth continuity of the muscle fibers shows the excellent accuracy of the tiling process. The connective tissue between different muscle groups is identified by white unstained regions. The thermal lesion is clearly visible as a light pink elliptical region with an approximate maximum diameter of 1.5 cm. The light pink staining indicates damaged muscle cells. The three lone crosses and the ten circles with an inscribed cross are the locations of fiducial needle tracks and ink fiducials, respectively. The location of these fiducials was accurately determined at high magnification from a live video window with a digitally imposed crosshair at its center. A snapshot of a live video window is shown for a typical fiducial needle track (b) and ink fiducial track (c). A fiducial needle track is easily identified by a significant break in the tissue with the surrounding muscle cells conforming to the shape of the needle, and by black ink inside the needle track. An ink fiducial is identified by a small break in the tissue with appropriately colored ink nearby. The crosshair is centered on each fiducial using a joystick controller. See text for more details.

Calibration ensured that a location on the live video window was appropriately located on the map image. The digitally imposed crosshair, at the center of the live video window, was centered over the corner of a small, recognizable object, and the stage coordinates were obtained. This procedure was repeated for each objective using the same object. This allowed us to determine an offset for each objective that was stored in a calibration file. In addition, the pixel size (gain) for each objective was stored in the lens calibration maintained by Scope-Pro. Thus, when the operator changed objectives to locate fiducials on a slide, the appropriate gain and offset were added to the stage coordinates to mark the correct position on the map image. This procedure ensured the center location of each objective was the same for the live video window but not for the eyepiece.

2.2.8 Validation of Registration for RF Ablation Experiments

We visually evaluated alignment of internal structures in MR, tissue, and histology images, using Regviz, a program written in IDL (Interactive Data Language, Research

System Inc., Boulder, CO) and created in our laboratory for visualizing and analyzing registered image volumes. This program had a linked cursor that allowed us to compare locations within multiple registered images. Using this program, we outlined various anatomical features in reference tissue images and copied them to corresponding images from histology and MR. It also allowed sectored displays where sections from different images were tiled together to show continuation of boundaries. Finally, the program allowed an overlay of a gray scale image from one type of acquisition to a color map

from another, called “hot iron” in IDL. These features allowed us to visually evaluate registration over aligned volumes of images.

We also quantitatively evaluated the registration accuracy using extra fiducial needle paths not optimized during alignment. There were two registration steps to align the MR and histology images (MR-tissue, tissue-histology). Errors in each step of the process as well as the complete error for the entire process were assessed. For these experiments, at least three needles were used: two for registration and one for evaluation. For every registered MR, tissue, and histology image, we manually localized the center of the extra needle path. For the MR images, we assumed needles were straight and obtained a least-squares fitted line from these points. For each point from histology images, we established a corresponding point for the fitted line in MR based on the closest 3D point along the fitted line. A 3D alignment error was determined by calculating the mean distance between corresponding points along the needle. We performed the same procedure to determine the 3D alignment error between MR and tissue images.

Although this method provides an independent measure for evaluating registration accuracy, there is error in localizing the fiducial needles in MR images. The localization error of needles in histology and tissue images was insignificant. To determine the effect of localization error of needles in MR, we estimated the error by calculating the variation from the line fitted to each needle in MR images, and compared this to the cumulative registration error.

2.3 RESULTS

2.3.1 Assessment of Registration

In vivo MR images were well registered with tissue and histology images. In Figure 2.4,

graphical overlays placed around needle tracks in MR images were copied to the tissue

images where they were well aligned with needle tracks in tissue and histology images.

Also, the intramuscular fat boundaries between muscle groups were well aligned in the

images. These fat boundaries were easily identified as unstained curves in histology, dark streaks in macroscopic tissue, and hyperintense streaks in MR images. The histological sample is smaller since we cut each thick-sectioned tissue slice to fit the histology-mounting block. In addition to these images, other adjacent slices aligned well indicating good 3D alignment.

We obtained the registration error of each alignment step (histology-tissue, tissue-

MR) for two in vivo rabbit thigh specimens. In Figure 2.5, we plotted, as a function of

each needle path, the mean 2D distance between points in histology and tissue images.

Over all needle paths for both tissue specimens, the result was 1.06 ± 0.30 mm (mean ±

SD). We assessed the tissue-MR registration error by calculating the mean 3D distance

between the extra needle paths not optimized during registration. In Figure 2.6, we

plotted the values as a function of needle registration pair, and the result over all needle

pairs was 1.04 ± 0.30 mm (mean ± SD). Similar results were obtained with needle paths

used for registration, with a distance of 0.89 ± 0.19 mm (mean ± SD).

Using the same registered image data as above, we determined entire 3D registration

error between MR and histology images. In Figure 2.7, we plotted, as a function of

needle registration pair, the mean 3D distance between the extra needle paths. Collapsing

over

Figure 2.4. Registration quality of in vivo MR, tissue, and histology images for adjacent slices in the two columns. Images are: T1-weighted with gadolinium contrast- enhancement MR (a,b), macroscopic tissue (c,d), and histology (e,f). Two needle tracks used for registration (circles) and two used for validation (squares) were marked in the in vivo MR images and copied to the tissue and histology images using RegViz. The needle tracks are dark regions surrounded by three bright spots in MR images, small dark regions in tissue photographs (arrows), and white holes in histology images (arrows). Occasionally, a needle track will be positioned in unstained connective tissue between different muscle groups and not be visible in a histology image (square in bottom left image without arrow). Excellent registration accuracy of MR, tissue, and histology images is clearly evident with good correspondence of needle tracks used for validation.

Figure 2.5. Plotted are mean histology-tissue registration errors as a function of the fiducial needle paths. Errors measured from specimen R105R (dark bars), and from specimen R105L (light bars), are both plotted. Specimens R105R and R105L had four and three needle paths, respectively. Each data point represents an average over typically five histological samples. The darker bar at the right (marked “AVG”) represent the average and standard deviation over all needle paths.

Figure 2.6. Plotted as a function of the needle path registration pair are average 3D MR-tissue registration errors for specimens R105R (dark bars) and R105L (light bars). Specimens R105R and R105L had six and three different needle registration pairs, respectively. Each data point represents an average over typically five tissue slices. The darker bar on the right (marked “AVG”) is the average and standard deviation over all needle registration pairs.

all needle registration pairs, the error was 1.32 ± 0.39 mm (mean ± SD), which compared favorably to the MR voxel dimensions (0.70 mm in plane and 3.0 mm thick). We obtained comparable results with needle paths optimized during registration, with a distance error of 1.27 ± 0.36 mm (mean ± SD). For a typical volume registration with eight histological samples, scatter plots of the x, y, and z displacements, as function of tissue slice, appeared random. The mean for each displacement was always less than

0.18 mm, a value close to zero. This is good evidence of random error rather than a systematic misregisration.

Figure 2.7. Plotted as a function of the needle path registration pair are mean 3D registration errors between MR and histology. The registrations are the same ones used for Figures 2.5 and 2.6. The average for each data point was over the same tissue samples as used in Figures 2.5 and 2.6. The average and standard deviation over all needle path registration pairs is the darker bar on the right (marked “AVG”).

To compare the registration error with the error associated with localizing needles in

MR images, we calculated the variation from the line fitted to each needle path. The localization error was 0.14 ± 0.05 mm (mean ± SD), approximately ten percent of the mean 3D registration error. Hence, the registration error reported is probably overestimated due to error in localizing the fiducial needles in MR images.

2.3.2 Comparison of Histology and MR Thermal Ablation Images

In Figure 2.8, careful examination of Masson trichrome histology, tissue, and in vivo MR images are shown. The central zone (zone 1) is light in the tissue images, hypointense in

T2-weighted MR images, and purple in histology due to damaged muscle cells.

Surrounding this is a transition zone (zone 2) that is reddish brown in tissue photographs, hyperintense in T2-weighted MR images, and both pink and red with significant extracellular space in histology. Zone 3 corresponds to normal tissue. Results were remarkably similar across several adjoining images. Data are from rabbit thigh specimen

RT3 that was sacrificed 45 minutes after ablation.

2.3.3 Comparison of Lesion Borders

In Figure 2.9, we copied a border for a region of necrosis identified manually in the

Masson trichrome histological samples to the registered tissue and T2-weighted MR images. This region was characterized by muscle cells with shrinkage or loss of nuclei, and contraction band necrosis (focal band-like coagulation of muscle contractile elements). This border matched features well in the tissue and MR images. The inner border of the dark brown region in the tissue images corresponded well to the histology

Figure 2.8. Zonal response of a typical thermal lesion. Registered images in the top row are: Masson trichrome stained histology (a), macroscopic tissue (b), and T2- weighted in vivo MR (c). The histological sample is smaller since we cut each thick-sectioned tissue slice to fit the histology-mounting block. Magnified (100X) Masson trichrome stained histology images (d-f) show characteristics of RF ablation in thigh muscle. Three labeled boxes were marked in the histology image and copied to the registered tissue and MR image. Box letters indicate the location of the magnified histology images. Zones of tissue damage are specified by box numbers. Excellent registration accuracy of histology and macroscopic tissue images is clearly evident with good correspondence of the outer tissue boundary in the histology image with the outer tissue boundary and a muscle group separation in the tissue image. The MR and tissue images are also well registered since the RF probe track (arrows) which is a bright spot in the MR image and a dark spot in the tissue image is closely aligned. In the histology image, the RF probe track is not visible since it was located in a tissue tear. border. The histology border was well aligned with the inner border of the hyperintense rim in MR images. To further analyze this, we compared boundaries marked by an observer in MR and histology images. For each tissue slice, an automatic algorithm determined equally spaced points, 0.25 mm apart, along a spline interpolated along the histology boundary. For each such point, the algorithm found the closest point along a spline interpolated from the corresponding MR boundary. A signed 2D Euclidean

distance between each point pair was calculated such that if the MR point is closer to the

boundary centroid than the histology point, the distance is negative, else it is positive.

This allowed us to establish if one boundary was interior or exterior to the other.

Preliminary results for rabbit thigh specimen RT3 showed the mean absolute distance between MR and histology boundaries was 1.17 mm. We determined that the MR boundary slightly overestimated the region of necrosis with a mean signed distance between borders of 0.85 mm. Results were remarkably similar in gadolinium contrast-enhanced T1-weighted MR images.

Figure 2.9. Comparison of tissue damage boundary identified in histology to tissue and MR image features. Images are from adjacent tissue slices in the two columns, and they include: histology images (a,b) with RF electrode track (arrow), tissue photographs from tissue slicing apparatus (c,d), and in vivo T2-weighted MR images (e,f). Boundary of region of cell necrosis identified previously are marked on the histology images with graphical overlays and copied to the registered tissue and MR images, where they match features in these images. Other white regions in the MR images are streaks of fat and the elliptical semitendinosus muscle group in the upper right corner. See text for details.

2.4 DISCUSSION

Our results suggest that it is possible to achieve 3D matching of histology and tomographic images such as MRI. Features such as the 2D warping registration of histology images and 3D needle registration of MR volumes are important steps for accurately aligning histology to in vivo MR images. For these experiments, we measured an overall registration error of 1.32 ± 0.39 mm (mean ± SD), which compares favorably to the MR voxel dimensions (0.70 mm in plane and 3.0 mm thick).

Unlike some potential alternatives, our method for estimating the 3D registration error can be practically applied. We can reliably localize the fiducial needle tracks in histology, tissue, and in vivo MR images. There are no 3D anatomical point landmarks and very few anatomical landmarks in our most common preparation, rabbit thigh muscle. Surface point fiducials attached to the thigh muscle will not be found in histological samples. We have not determined how to add internal point fiducials that can be reliably localized in tissue and histological specimens; they can be dislodged during tissue slicing or missed between thick tissue slices. With the fiducial needles, we expose the tips of the needles at the tissue block face and highlight them with ink for easy identification in the tissue photograph. The needles are withdrawn as we step through the tissue and slice it.

The fiducial needle error measure is sound. Because it is a local measure, it is comparable to the measure of most interest: the displacement of thermal ablation boundaries. By adding additional needles not used in the registration, an independent measure of registration quality is obtained. It is understood that if the needles are placed parallel to one another, there is no constraint for registration along the z-axis, and the

error measure is meaningless. Previously, we performed many simulation experiments

and determined that the method was quite accurate as long as the angle between the needles exceeded 15 deg, even with a needle localization error (11). We always use at least 30 deg. In simulations with very large needle localization errors, we determined that the 3D displacements between needle tracks reliably provided a diagnostic for the true registration error determined in our simulations. The biggest problem with using needle track displacement as a measure of error is that it also includes needle localization uncertainty. For histology and tissue images, the uncertainty was probably negligible.

However, for MR images, Lewin et al. (13) showed that the error in localizing a needle with our MR system and scanning sequence was typically within 1.0 mm. Hence, as much as 1.0 mm of our 1.3 mm average registration error might be due to needle track localization error.

There are various potential experimental approaches for spatially correlating MR and tissue response. First, biopsy samples can be obtained under MR guidance. However, such samples cannot finely sample the lesion and vicinity as required for a complete

tissue damage evaluation. Second, rapid freezing of whole tissue sections might be an

alternative approach to chemical methods of fixation. Cryosectioning was demonstrated

for the visible human project (14) and for sectioning excised human and animal brains

(15-17). A disadvantage of freezing is that cellular morphology is degraded. Moreover,

cryosectioning does not necessarily eliminate all geometric problems as many open

cavities were created during the sectioning of the visible humans. Third, there are various

2D methods for correlating MR and tissue response. For these 2D approaches, the

experiment is lost with no possible recovery when the plane of interest is missed in

cutting. Our 3D technique for relating histology to MR images addresses these

limitations.

We overcame several challenges to make our 3D registration method work reliably.

During our initial tissue dissections, it was not easy to maintain the orientation of the fiducial needles. Also, poorly positioned lesions made dissection extremely difficult.

During tissue slicing, movement and tearing of the tissue caused warping of the tissue.

Two of our preliminary rabbit thigh experiments gave a third needle error value greater than 3 mm. Further visual investigation revealed that tissue deformation, presumably during dissection and tissue slicing, caused the fiducial needles to change their relative orientation. A large distance metric was diagnostic of this scenario. These earlier rabbit experiments were dropped from further analysis. The problems identified above have been largely minimized with experience. Of the last 20 experiments, 90 percent were successful; failures were due to movement of a fiducial needle during tissue handling.

The results presented here are only those obtained during experiments in which we had

extra fiducial needles to assess accuracy.

Another potential limitation is tissue distortion due to shrinkage and swelling. By

using a uniform scale parameter in the needle registration, we accounted for tissue

shrinkage due to fixation. For our thigh muscle experiments, shrinkage was typically five

percent, a value consistent with previous studies (18), which showed a mean shrinkage of

2.3 percent for muscle tissue. A potential source of local distortion might occur with any

tissue swelling due to the hyperacute inflammatory response. However, such swelling

should occur before insertion of the needle fiducials, and any regional distortion should

be consistent across the image data.

From histology images, there was a distinct central region characterized by muscle cells with shrunken nuclei and contraction band necrosis. Surrounding this was an exceptionally well demarcated region having increased extracellular space, and cells with damaged muscle contractile proteins. This tissue response was very sharply delimited against adjacent normal tissue. In addition, the border of the central region correlated with color changes from purple to red as seen in Masson trichrome stained tissue. It is believed that the central region corresponds to the area of irreversible cell damage (19).

The surrounding region with badly damaged cells may go onto cell death. This will be addressed in future survival experiments.

A preliminary comparison of MR and tissue response following RF thermal ablation showed that the region within the inner border of the hyperintense rim in MR images closely corresponds to the necrotic region, and the surrounding region with damaged cells may become necrotic. These results are consistent with previous studies that investigated the correlation of MRI and histology following focused ultrasound (7) and laser ablations

(5). These studies suggest that the region within the inner border of the hyperintense rim in MR images corresponds to coagulation necrosis, and the region between the inner and outer border of the hyperintense rim will possibly become necrotic.

We conclude that our 3D methodology can be used to accurately map tissue response to MR thermal lesion images. Preliminary results show that in the rabbit thigh muscle, the inner and outer border of the hyperintense rim in MR images corresponds to the border of complete and partial irreversible cell damage. This is good evidence that iMRI thermal lesion images can be used for real-time feedback during thermal RF ablation treatments.

WORKS CITED

1. Jolesz, F. A., Bleier, A. R., Jakab, P., Ruenzel, P. W., Huttl, K., and Jako, G. J. MR Imaging of Laser-Tissue Interactions. Radiology 1988;168(1):249-53.

2. McDannold, Hynynen, K., Wolf, D., Wolf, G., and Jolesz, F. MRI Evaluation of Thermal Ablation of Tumors With Focused Ultrasound. J.Magn Reson.Imaging 1998;8(1):91-100.

3. Merkle, E. M., Haaga, J. R., Duerk, J. L., Jacobs, G. H., Brambs, H. J., and Lewin, J. S. MR Imaging-Guided Radio-Frequency Thermal Ablation in the Pancreas in a Porcine Model With a Modified Clinical C-Arm System. Radiology 1999;213(2):461-7.

4. Boaz, T. L., Lewin, J. S., Chung, Y. C., Duerk, J. L., Clampitt, M. E., and Haaga, J. R. MR Monitoring of MR-Guided Radiofrequency Thermal Ablation of Normal Liver in an Animal Model. J.Magn Reson.Imaging 1998;8(1):64-9.

5. Morrison, P. R., Jolesz, F. A., Charous, D., Mulkern, R. V., Hushek, S. G., Margolis, R., and Fried, M. P. MRI of Laser-Induced Interstitial Thermal Injury in an in Vivo Animal Liver Model With Histologic Correlation. J.Magn Reson.Imaging 1998;8(1):57-63.

6. Lewin, J. S., Connell, C. F., Duerk, J. L., Chung, Y. C., Clampitt, M. E., Spisak, J., Gazelle, G. S., and Haaga, J. R. Interactive MRI-Guided Radiofrequency Interstitial Thermal Ablation of Abdominal Tumors: Clinical Trial for Evaluation of Safety and Feasibility. J.Magn Reson.Imaging 1998;8(1):40-7.

7. Chen, L., Bouley, D. M., Harris, B. T., and Butts, K. MRI Study of Immediate Cell Viability in Focused Ultrasound Lesions in the Rabbit Brain. J.Magn Reson.Imaging 2001;13(1):23-30.

8. Lee, Z., Sodee, D. B, Faulhaber, P. F., Lancaster, T. L, MacLennan, G. T., and Wilson, D. L. Comparison of SPECT and PET Imaging for Prostate Cancer With Histological Correlation. Journal of Nuclear Medicine 2001;42:294 (Abstract).

9. Salem, K. A., Szymanski-Exner, A., Lazebnik, R. S, Breen, M. S., Gao, J., and Wilson, D. L. X-Ray Computed Tomography Methods for in Vivo Evaluation of Local Drug Release Systems. IEEE Transactions on Medical Imaging 2002;(In Press).

10. Besl, P. J. and McKay, H. D. A Method for Registration of 3-D Shapes. IEEE Transactions on Pattern Analysis and Machine Intelligence 1992;14(2):239-56.

11. Lazebnik, R. S., Lancaster, T. L., Breen, M. S., Lewin, J. S., and Wilson, D. L. Volume Registration Using Needle Paths and Point Landmarks for Evaluation of Interventional MRI Treatments. IEEE Trans.Med.Imaging 2003;22(5):653-60.

12. Lancaster, T. L and Wilson, D. L. Correcting Spatial Distortion of Histological Images. Annals of Biomedical Engineering 2002;(Submitted).

13. Lewin, J. S., Duerk, J. L., Jain, V. R., Petersilge, C. A., Chao, C. P., and Haaga, J. R. Needle Localization in MR-Guided Biopsy and Aspiration: Effects of Field Strength, Sequence Design, and Magnetic Field Orientation. AJR Am.J.Roentgenol. 1996;166(6):1337-45.

14. Spitzer, V. M and Whitlock, D. G. Atlas of the Visible Human Male: Reverse Engineering of the Human Body. Jones and Bartlett 1998.

15. Toga, A. W., Goldkorn, A., Ambach, K., Chao, K., Quinn, B. C., and Yao, P. Postmortem Cryosectioning As an Anatomic Reference for Human Brain Mapping. Comput.Med.Imaging Graph. 1997;21(2):131-41.

16. Toga, A. W., Ambach, K., Quinn, B., Hutchin, M., and Burton, J. S. Postmortem Anatomy From Cryosectioned Whole Human Brain. J.Neurosci.Methods 1994;54(2):239-52.

17. Toga, A. W, Quinn, B., and Ambach, K. A High Resolution Digital Collection of Cryosectioned Whole Brain. J.Cerebal Blood Flow and Metabolism 1993;13(1):820.

18. Cutts, A. Shrinkage of Muscle Fibres During the Fixation of Cadaveric Tissue. J.Anat. 1988;160:75-8.

19. Baroldi, G. Myocardial Cell Death, Including Ischemic Heart Disease and its Complications. Cardiovascular Pathology, New York, Churchill Livingstone. 3 ed.2001. pp.202-6.

CHAPTER 3

Radio Frequency Thermal Ablation: Correlation of Hyperacute MR Lesion Images with Tissue Response

3.1 INTRODUCTION

Solid tumors and other pathologies can be treated using radio frequency (RF) thermal ablation under interventional magnetic resonance (MR) image guidance (1-5). MR imaging (MRI) has several advantages including the lack of ionizing radiation, excellent soft tissue discrimination, sensitivity to blood flow and temperature, and ability to image at any angle (6;7). In addition, MR imaging appears to be superior to ultrasound or x-ray computed tomography for monitoring the effects of RF thermal ablation treatments (8;9).

The major contribution of MRI is its potential to monitor the zone of thermal tissue destruction during the procedure. To monitor an ablation procedure, MRI can intermittently acquire temperature images during heating and structural lesion images during and after heating. We investigated this unique ability to monitor treatment by comparing hyperacute (several minutes post-ablation) MR lesion images to tissue damage as seen histologically. The ability of these MR images to accurately predict of the region of eventual cell death would allow complete destruction of the disease tissue volume, including margin, while avoiding the destruction of nearby normal tissue of critical importance.

If one can prove that the hyperacute MR signal accurately reflects tissue damage and destruction, then MR can be used to confidently monitor treatment during a procedure.

This will enable one to adjust the size and shape of the treated region, through additional heating or displacement of the RF electrode to a new location, so as to ensure coverage of the pathology. The ability to monitor therapy will enable one to extend the method to the safe destruction of tumor adjacent to vital structures that might be damaged by heating such as the gall bladder, bowel, and especially the brain, where collateral damage must be minimized.

Our careful correlation of hyperacute MR lesion images with the corresponding tissue response has several advantages. First, this could allow for a better understanding of mechanism responsible for the MR signal changes, particularly the hypointense center and hyperintense rim. These MR signals changes could be due to different zones of tissue damage or different zones of increased extracellular space. Second, our analysis may allow for a more complete understanding of the tissue response at the margin of the lesion. Third, this would allow one to compare the mechanism and pattern of tissue destruction of RF ablation to other ablation techniques such as laser and focused ultrasound, where similar MR lesion appearances are observed. Finally, this correlation could better explain the macroscopic tissue color changes, which were used by others to assess tissue damage (10-14).

3.2 MATERIALS AND METHODS

3.2.1 Methods for RF Ablation Experiments

The comparison of in vivo MR thermal lesion images to the tissue response required both experimental and registration methods as described in more detail below. Briefly, experimental methods included RF thermal ablation of a rabbit thigh model, post-ablation

MR imaging, tissue slicing and photographing, and histological processing and imaging.

It was extremely important to minimize tissue deformation and destruction during the dissection and slicing processes for accurate registration. Using the tissue photographs as a reference, we aligned histology and MR images with two registration methods. First, the MR volume was aligned to the volume of tissue photographs by performing a 3D fiducial needle registration. Second, we aligned the histology images with the tissue photographs using a 2D warping registration. To correlate MR images with the tissue response, we segmented tissue damage boundaries in histology images, and compared them to segmented lesion boundaries in registered MR images.

3.2.2 RF Ablation in Rabbit Thigh Model and MR Imaging

Following a protocol approved by the Institutional Animal Care and Use Committee, we anesthetized five New Zealand White rabbits (3.0-3.5 kg) with a 2 ml IM injection of a cocktail (0.6 ml/kg) – a combination of Ketamine (0.26 ml/kg), Xylazine (0.26 ml/kg), and Acepromazine (0.08 ml/kg) (each manufactured by Phoenix Scientific, St. Joseph,

MO), and maintained the anesthesia with IM injections every 20-40 minutes that alternated between 0.5 ml of ketamine (0.15 ml/kg) and 1 ml of cocktail (0.3 ml/kg).

After shaving each animal’s abdomen, and left thigh, we placed the rabbits in the prone position within the gantry of a clinical 0.2 T C-arm MR imaging system (Siemens

MAGNETOM OPEN, Erlangen, Germany). The legs of each rabbit were secured to a customized Plexiglas support, which prevented movement of thighs. After positioning two 8 x 12 cm wire mesh grounding pads coated with conductive gel (Aquasonic 100:

Parker Laboratories, Orange, NJ) on each rabbit’s abdomen, we placed the thighs within

a 12 cm diameter multi-turn solenoid, receive-only coil.

Under MR guidance, we inserted an MR-compatible 17-G titanium RF electrode with

a 10 mm exposed tip (Radionics, Burlington, MA) percutaneously into the thigh muscle.

Prior to ablation, we acquired an MRI volume of the thigh. A T2-weighted

turbo-spin-echo sequence was applied with TR/TE/NSA parameters of 3362/68/8 that

gives 256 x 256 x 9 voxels over a 180 x 180 x 27-mm FOV to yield 0.70 x 0.70 x 3.0-mm

voxels oriented to give the highest resolution for slices approximately perpendicular to

the RF electrode. Lesion formation was achieved by increasing the local tissue

temperature with resistive heating by delivering RF electric current between the electrode

tip and ground pads. We applied RF energy for two minutes using a 100 W RF generator

operating at 500 kHz (RFG-3C, Radionics, Burlington, MA). The tip of the RF electrode

was maintained at a temperature of 90 ± 2°C using a thermistor within the electrode tip.

Immediately after ablation, a 22-G x 1 inch IV catheter (Terumo Medical Corporation,

Elkton, MD) was inserted in a dorsal ear vein in each rabbit. At least two MR-compatible

22-G, 10 cm fiducial needles (E-Z-EM, Westbury, NY) were inserted into the thigh near

the thermal lesion with one fiducial approximately parallel and the other fiducial

approximately at an angle of 45° to the RF electrode. The RF electrode was removed from the thigh before imaging to prevent MR image artifacts.

Approximately 10 minutes after ablation, we acquired MRI volumes of the thigh. A

T2-weighted turbo-spin-echo sequence was applied with the same parameters and orientation as the pre-ablation image acquisition described above. We also applied a

T1-weighted spin-echo sequence with TR/TE/NSA parameters of 624/26/6 that gives

256 x 256 x 9 voxels over a 180 x 180 x 27 mm FOV to yield 0.70 x 0.70 x 3.0 mm voxels oriented the same as the T2-weighted sequence. At least 5 minutes before acquiring the T1-weighted spin-echo images, we administered an IV injection of gadolinium contrast medium (0.2 mmol/kg, gadopentate dimeglumine: Berlex

Laboratories, Wayne, NJ). The rabbits were sacrificed approximately 45 minutes after the ablation using a barbiturate overdose technique via intravenous administration of

0.1 ml/lb of pentobarbital sodium (Euthasol, 390mg/ml, Diamond Animal Health Inc.,

Des Moines, IA). After fixation, we removed the thigh from the bone for gross pathological and histological examination.

3.2.3 Tissue Slicing and Acquisition of Calibrated Tissue Images

We obtained 3 mm tissue slices by using a specially designed apparatus that included a tissue platform, digital camera (DSC-D770, Sony, Japan), and a linear displacement device (Rack and Pinion Slide, Edmund Scientific, Barrington, NJ) for accurate stepping of the platform in small increments. To orient the specimen and reduce deformation during slicing, we embedded the specimen in tissue embedding wax (50-54°C Parablast

X-tra, Oxford Labware, St. Louis, MO). By orienting both planes perpendicular to a fiducial needle, we ensured the tissue-slicing plane and the plane of the MR image slices were approximately parallel. This minimizes image blurring of the re-sliced images when there is out-of-plane tilting with respect to the thick MR image slices. We sliced the specimen with a 12.8 inch autopsy knife (Tissue-Tek Accu-Edge Semi-Disposable

Autopsy Knife System, Sakura Finetek, Japan) using vertical supports of the apparatus as a guide. We photographed the tissue block face, advanced the platform by 3 mm, and

sliced the tissue, repeating the process until the specimen was traversed. The tips of the fiducial needles were exposed at the tissue block face for easy identification in each photograph, and stepped back slightly beyond the plane of the next tissue slice. We calibrated these macroscopic tissue images using a ruler in the plane of the tissue slice.

The tissue images yield square pixels that were typically 0.17 mm on a side. We embedded the tissue slices in paraffin and obtained hematoxylin and eosin (H&E) and

Masson trichrome (MT) stained histological sections mounted on glass microscope slides.

3.2.4 Acquisition of Histology Images

Histology slides were digitized using a video microcopy system that consisted of a light microscope (BX60, Olympus, Japan), video camera (DXC-390, Sony, Japan), position encoded motorized stage (ProScan, Prior Scientific, Rockland, MA), and controller software (Image-Pro with Scope-Pro, Media Cybernetics, Silver Spring, MD). To obtain an image of the entire slide, we used a software function that drove the motorized stage, acquired a series of photographs, and seamlessly combined the photographs to form one large tiled image with pixels that were typically 5.21 um on a side. By reducing the tiled image to 10 percent of its original size, we created a smaller map image on which we marked the locations of tissue damage boundaries.

3.2.5 Segmentation of Tissue Damage Boundaries in Histology

To compare the thermal lesion appearance in MR images with the tissue response, we manually segmented (marked) boundaries of tissue damage identified in histology. Under

the supervision of a pathologist (author MF), we established criteria to identify cell damage. Tissue damage was based on changes in cell morphology and stain color on

H&E and MT stained sections. In addition, we evaluated cell damage based on the loss of muscle’s naturally occurring birefringence, which has been shown to be a reliable indicator of irreversible cell damage (15;16). Birefringence properties were determined in

MT stained sections under polarized light. Both an inner and outer tissue damage boundary was marked for each MT stained section. The inner boundary separated a central region (zone H1) characterized by muscle cells with shrunken nuclei, contraction band necrosis/coagulative myocytolysis, and complete loss of birefringence from a well demarcated region (zone H2) with similar histological changes but only a partial loss of birefringence. We also marked an outer tissue damage boundary that divided a conspicuous region (zone H3) characterized by cells that appeared frankly necrotic, were shrunken and fragmented, and were associated with distinct interstitial edema. Zone H3 was very sharply delimited from adjacent normal tissue (zone H4). Hence, we marked the inner and outer boundary of a region occupied by the union of two zones (zone H2+H3).

A single observer manually marked the two tissue damage boundaries using a previously developed video microscopy system and a software-scripting program written in Image-Pro and Scope-Pro (17). The observer had experience examining histology without knowledge of MR lesion boundaries. With a digitally imposed crosshair at the center of a live video window, the operator panned around the slide under joystick control and identified boundaries of interest. While locating boundaries, the operator could switch microscope objectives to acquire images at higher or lower magnification levels. On each boundary, we identified 15-30 points by centering the crosshair over each

point of interest and clicking a graphical user interface button to acquire the stage

coordinates. Each boundary point was marked at the appropriate location on the map

image with a colored graphical overlay. The end result of this marking process was a map

image with two tissue damage boundaries that can be compared with lesion boundaries seen in MR images.

3.2.6 Registration of MR and Histology Images

A previously developed three dimensional (3D) registration method was used to align the histological and in vivo MR image data (17). Briefly, we used the macroscopic tissue images as the reference and registered histology and MR images to them with two different computer alignment steps. First, the MR volume was aligned to the volume of tissue images by registering the fiducial needles placed near the lesion (18). Second, we registered the histology images with the tissue images using a two dimensional (2D) warping technique that aligned internal features and the outside boundary of histology and tissue images (19). The above steps allowed us to match a pixel in a histology image with an interpolated sample from the in vivo MR image volume. This registration method was previously validated (17), and the accuracy determined from displacement of needle fiducials was estimated to be 1.32 mm ± 0.39 mm (mean ± SD).

3.2.7 Segmentation of Lesion Boundaries in MR Images

To compare the histologic tissue damage boundaries with lesion appearance in MR, we manually segmented MR lesion images. Both the inner and outer boundaries of the lesion’s hyperintense rim were segmented for each CE T1 and T2-weighted MR image

volume. The investigators established criteria for boundaries under the supervision of a radiologist specializing in interventional MRI RF ablation. Prior to actual segmentation, observers segmented a training set consisting of images similar to the experimental data.

Training set results were compared among all investigators and further training performed until consistent results were obtained.

Two of the authors independently performed the segmentation using a freehand region of interest (ROI) tool in Analyze 3.1 (Analyze Direct, Lenexa, Kansas) image analysis software. These observers had experience looking at MR images without knowledge of histology boundaries. Important software features included the ability to look at adjacent

MR image slices simultaneously and a region of interest (ROI) editing tool. The observers followed a strict protocol. All segmentation was performed on the same workstation in a darkened room. We perceptually linearized the display using Optical

(ColorVision Optical, Rochester, NY) and validated with a step wedge image. By adjusting the vertical and horizontal controls of the display, we obtained square pixels with 0.2 mm edge length. Window and level settings for each image set were fixed prior to segmentation such that the qualitative contrast between the inner and outer lesion zones was maximized. The T2-weighted MR images acquired prior to ablation could be viewed simultaneously on our display and were utilized to visualize background features such as fat and connective tissue so as to appropriately eliminate these structures from the segmented lesions. Each observer was blinded from the results of the other observer. To minimize bias, T2 and CE T1-weighted images for a given rabbit were never segmented during the same day.

Following segmentation, the ROI boundary coordinates were exported and analyzed using MATLAB 6.1 (Mathworks, Natick, MA). Each segmented point was assigned a 3D coordinate, in millimeter dimensions, based on the voxel dimensions. The set of boundaries was utilized for subsequent comparison with tissue damage boundaries from histology.

3.2.8 Boundary Comparison of MR to Tissue Response

To evaluate the ability of MR images to predict tissue response, we used software programs written in MATLAB to directly compare the manually segmented boundaries from MR and histology images. Each boundary coordinate was assigned a 2D coordinate, in millimeters, based on the pixel size. For each tissue slice, an automatic algorithm determined equally spaced points, 0.25 mm apart, along a spline interpolated along the histology boundary. For each such point, the algorithm found the closest point along a continuous spline interpolated from the corresponding MR boundary. A signed 2D

Euclidean distance between each point pair was determined such that if the MR point is closer to center of lesion than the histology point, the distance is negative, else it is positive. This allowed us to determine if one boundary was interior or exterior to the other.

To examine the effect of MR segmentation error on any discrepancy between tissue damage and MR lesion boundaries, we compared the interobserver variability of the two observers’ MR boundaries to the difference between MR and histology boundaries. The

MR and histology boundaries were used to calculate the Williams index (WI), a ratio of the average absolute distance between multi-observer MR boundaries to the average

absolute distance between the multi-observer MR boundaries and histology boundaries (20). A WI of 1.0 implies the difference between histology and MR boundaries is not more than the disagreement among the MR observers themselves.

Statistical analysis was performed with the two-sample t-test. A P-value less than

0.05 was considered to indicate a statistically significant difference.

3.3 RESULTS

In Figure 3.1, we show typical in vivo MR lesion images acquired approximately

45 minutes post-ablation. The plane of the MR image was oriented approximately perpendicular to the RF electrode. These MR images reveal a significant change of the

MR signal in the vicinity of the RF current source. The characteristic elliptical appearance for both T2 and CE T1-weighted MR images has a central core, hereafter called zone M1, surrounded by a hyperintense margin (zone M2). Beyond zone M2, the

MR signal is isointense to surrounding muscle (zone M3). Other experiments gave similar results.

In Figure 3.2, histology images of a thigh muscle from a rabbit sacrificed approximately 45 minutes post-ablation are shown. The histological samples were obtained approximately perpendicular to the RF needle. Surrounding the RF needle track, a distinct thermal lesion can be seen. In Figure 3.2, rows H1, H2, H3, and H4 show images of four distinct histological zones.

In the central region of the lesion (zone H1), the architecture of the skeletal muscle appears intact, with a size, shape, and distribution of cells similar to normal skeletal muscle. However, in the majority of cells, the cell nuclei are smaller and somewhat

Figure 3.1. A typical CE T1-weighted (left) and T2-weighted (right) in vivo MR image acquired minutes after ablation. The thermal lesion (vertical arrow) is the elliptical region that has an isointense/hypointense central core surrounded by a hyperintense rim in both images. The small bright dot at the center of the lesion is blood and/or interstitial fluid that filled the track of the RF electrode, which was withdrawn prior to imaging. The small dark region surrounded with bright spots (horizontal arrow) is an MR image artifact of a fiducial needle, which was used to align the MR and tissue image volumes.

pyknotic, and the cytoplasm shows evidence of contraction band necrosis (focal band-like

coagulation of contractile elements) or coagulative myocytolysis (granular myofibrillar

degeneration). The cytoplasm of the skeletal muscle cells is eosinophilic (pink) like that

of normal skeletal muscle cells on H&E stain, but is metachromatic rather than red on

MT stain. Often the ratio of cells with coagulative myocytolysis to those with contraction

band necrosis increased with increasing distance from the RF needle track. Under

polarized light, the cells are dark due to significant loss of birefringence.

Further from the lesion center, there is a well demarcated transition zone (zone H2)

with both metachromatic and red cells on MT stain, and more extracellular space than

normal muscle. The cells show evidence of contraction band necrosis or coagulative

myocytolysis. Under polarized light, there are both dark and bright muscle cells due to a

partial loss of birefringence.

Surrounding zone H2, there is distinct region (zone H3) that is significantly paler than

the zone H1 and H2 on both H&E and MT stains. However, in many cases, the skeletal

muscle cells themselves are more hypereosinophilic (darker pink/red on H&E and MT

stains) than normal skeletal muscle cells. These hypereosinophilic cells appear necrotic,

lack nuclei, are shrunken and distorted with a wavy or fragmented appearance, and are

associated with marked interstitial edema. We sometimes found pools of red blood cells

indicative of hemorrhage. Under polarized light, there are both dark and bright cells due

to a partial loss of birefringence. No inflammatory cells are seen in either zone H1, H2, or

H3. Blood vessels appear intact in all three zones. It should be noted that zone H2+H3

(union of zone H2 and H3) has more extracellular space than normal muscle and partial

loss of birefringence.

This tissue response is very sharply delimited against adjacent normal muscle tissue

(zone H4). The normal cells appear bright under polarized light since they are completely birefringent. Results were remarkably similar in other experiments.

In Figure 3.3, we copied the inner and outer boundary of H2+H3 identified manually in MT stained histological samples to the registered T2 and CE T1-weighted MR images.

These histology boundaries matched features well in the MR images. Zone H2+H3 was well aligned with the hyperintense rim in both T2 and CE T1-weighted MR images.

Results were remarkably similar across several adjoining tissue slices.

Figure 3.2. Typical histology images show a hyperacute zonal response. Images are: H&E (left) and MT stain with unpolarized light (middle), and MT with polarized light (right). Images of the entire histology slide (top row) were formed by tiling a series of pictures as described in Methods. Magnified (100X) histology images (rows H1-H4) show characteristics following RF ablation in rabbit thigh muscle. Boxes in the images show the location of the magnified histology images. A distinct therm al lesion can be seen as an elliptical region stained light pink with H&E (top left), light purple with MT (top middle), and darkened with polarized light (top right). In the central zone of the lesion (zone H1), the RF needle track is seen as an empty cavity. See text for details.

Figure 3.3. Comparison of tissue damage boundaries identified in histology with T2 and CE T1-weighted MR image features. The images in the left and right column are identical except for graphical overlays on the left showing the inner (green) and outer (yellow) boundary of zone H2+H3 for tissue slice R16-14. The registered images are: T2-weighted (top) and CE T1-weighted (middle) MR and MT stained histology images (bottom). Borders are marked on the histology image with graphical overlays and copied to MR images, where they match features in these images. Other bright regions in the MR images are streaks of fat and the elliptical femur bone in the lower left corner.

To further analyze this, we quantitatively compared the boundaries of M2 with

H2+H3. In Figure 3.4, the 2D signed distance for the inner and outer boundaries was plotted as a function of distance along the boundary of H2+H3. We used the boundary of

H2+H3 as the reference and measured distances to corresponding points along the

boundary of M2. Results are shown for M2 boundaries marked by two observers.

Collapsing over both tissue slices and observers for the inner and outer boundaries respectively, the mean absolute distance was 1.06 ± 0.34 mm (mean ± SD) and

0.51 ± 0.08 mm for T2-weighted MR images, and 0.77 ± 0.14 mm and 0.47 ± 0.21 mm for CE T1-weighted MR images. These values compare favorably to the in-plane MR voxel dimension (0.70 mm) and slice thickness (3.00 mm). This is good evidence that the boundaries of M2 for T2 and CE T1-weighted MR images can accurately predict the boundaries of H2+H3.

Using registered image data from all five lesions, we determined the discrepancy between the boundaries of M2 and H2+H3. In Figure 3.5, we plotted the mean signed and absolute distance for inner and outer boundaries as a function of the MR observers.

Collapsing over both MR observer boundaries for the inner and outer boundaries respectively, the mean signed distance was not significantly different from zero (P=0.456,

0.142) for T2 and (0.326, 0.053) for CE T1-weighted MR images, indicating insufficient bias between corresponding boundaries. For the inner and outer boundaries respectively, the mean absolute distance was 1.04 ± 0.30 mm (mean ± SD) and 0.96 ± 0.34 mm for T2 and 1.00 ± 0.34 mm and 0.94 ± 0.44 mm for CE T1-weighted images, which compares favorably to the in-plane MR voxel dimension (0.70 mm) and slice thickness (3.00 mm).

Averaging the two M2 observer distances for each slice, we determined that the mean absolute distances for T2-weighted MR images were not significantly different from those for CE T1-weighted images (P=0.745, 0.818) for the inner and outer boundaries, respectively.

Figure 3.4. Plot of signed distance between inner (left) and outer (right) H2+H3 and M2 boundaries for T2 (top) and CE T1-weighted (bottom) MR images as a function of distance along the histology boundary. The adjacent tissue slices are R16-13 (thin line) and R16-14 (thick line). The red and blue lines represent the M2 boundary marked by observer 1 and 2, respectively. Signed distance is negative when MR boundary is closer to center of lesion than histology boundary, else distance is positive.

To examine the effect of MR segmentation error on the small discrepancy between histology and MR boundaries, we compared the interobserver variability of the two observers’ M2 boundaries to the difference between M2 and H2+H3 boundaries. We calculated the Williams index (WI), ratio of the average absolute distance between the two observers’ M2 boundaries to the average absolute distance between the M2 and

H2+H3 boundaries. In Table 3.1, we show the average absolute distance between H2+H3 and two observers’ M2 boundaries (HOD), average absolute distance between the two observers’ M2 boundaries (IOD), Williams index (WI), and 95% confidence interval

(CI). A CI with an upper limit greater than 1.0 indicates that there is as much variability between H2+H3 and M2 boundaries as there is between the two observers’ M2 boundaries (20). Since each upper limit of the CI was less than 1.0, there is greater agreement between the observers’ M2 boundaries than between H2+H3 and M2 boundaries. Upper limit CI values between 0.57 to 0.80 indicate that the variability between H2+H3 and M2 boundaries is about twice the variability of the observers’ M2 boundaries. This is good evidence that the MR boundary segmentation error accounts for approximately one-half of the discrepancy between MR and histology boundaries.

Figure 3.5. Plot of mean absolute (top) and signed (bottom) distances between inner (left) and outer (right) H2+H3 and M2 boundaries for T2 and CE T1-weighted MR im ages. The dark and light bars are the M2 boundaries segmented by observer 1 and 2, respectively. Each bar represents the average and standard deviation over 12 and 14 tissue slices for the inner and outer boundary, respectively. Note that two additional tissue slices that contained only an outer boundary were obtained from the lesion ends. Signed distance is negative when MR boundary is closer to center of lesion than histology boundary, else distance is positive.

3.4 DISCUSSION

Our results suggest that it is possible for hyperacute T2 and CE T1-weighted MR lesion images to predict the tissue response to RF ablation. That is, we determined that zone M2 from MR and the region occupied by the union of two zones seen in histology, H2+H3, matched very well. Other regions, M1 and H1, and M3 and H4, also necessarily correspond. Features of our method such as 3D registration of in vivo MR images to histology images, accurate segmentation of tissue damage boundaries on tiled images of large-format histology slides, and reliable assays to determine tissue damage such as polarized light assessment of muscle birefringence, are important steps to accurately correlate the tissue response to in vivo MR thermal lesions images.

3.4.1 Tissue Response to RF Ablation

It is believed that histology zones H1, H2, and H3 correspond to the region of eventual cell death. Although the exact determination of cell viability is challenging, the presence of contraction band necrosis and coagulative myocytolysis, and a loss of birefringence suggest that the cells within zones H1 and H2 are non-viable. Zone H3 is clearly

non-viable due to the fragmentation of the cells. Upon thermal injury, a loss of

birefringence occurs due to the disarray of the regular matrix of the actin and myosin

molecules, and this has been previously shown to correspond to a region of eventual

necrosis (15;16). We recognize that the birefringence assay for cellular viability is unique

to muscle tissue. In addition, although necrotic changes specific to muscle were reported,

the MR thermal lesion appearance and resulting general coagulative changes due to

thermal treatment are common to many tissues (21-23). We are also performing studies

with animals sacrificed several days post-ablation, which should unequivocally reveal the

complete extent of tissue necrosis in histology.

It appears likely that the contraction band necrosis and coagulative myocytolysis we

describe in RF ablation of skeletal muscle is due to muscular hypercontraction in

response to an influx of calcium ions as a result of thermal injury, rather than to direct

thermal coagulation. As hypercontracted skeletal muscle cells may be more rigid than their normal counterparts, it is also possible that the cellular distortion/fragmentation and edema seen in the zone H3 of the lesions may be due to mechanical stress from the surrounding normal skeletal muscle cells.

3.4.2 Assessment of Correlation between MR and Histology Boundaries

It is possible that the MR and histology boundaries match exactly and the small differences are less than our ability to measure for a variety of reasons. First, the mean absolute differences (typically 1.0 mm) compare favorably to the MR in-plane voxel

width (0.70 mm) and thickness (3.00 mm). Probably because of the partial volume effect,

there is uncertainty of at least one voxel width as to where to visually place the edge. This

manual MR segmentation uncertainty must limit the study resolution. The WI values are consistent with this observation; they indicate that the variability between histology and

MR boundaries is about twice the variability of the two observers’ MR boundaries. Thus, the MR segmentation error possible accounts for approximately one-half of the discrepancy between MR and histology boundaries.

Second, registration errors arise from both the alignment of the MR volume with the tissue sections and the warping of histological sections to the tissue sections. By comparing needle fiducial locations, it was previously determined that the error of the entire registration procedure is random, not systematic, across specimens (17). The registration method work reliably after several challenges, such as warping of the tissue due to movement and tearing of the tissue during dissection and tissue slicing, were overcome with experience. The 3D registration method also compensated for slight tissue shrinking due to fixation by using a uniform scale parameter in the needle registration.

For these thigh muscle experiments, shrinkage was typically five percent, a value consistent with previous studies (24), which showed a mean shrinkage of 2.3 percent for muscle tissue. In addition, a potential source of local distortion might occur with any tissue swelling due to edema. However, such swelling should occur before insertion of the needle fiducials, and any regional distortion should be consistent across the image data. Although we have done much to limit registration error, it must account for a considerable amount of the discrepancy between boundaries.

Third, we note that although registration was performed in 3D, boundary distance calculations were performed on a 2D basis due to the slice nature of histological sections.

This approach tends to overestimate distances since the closest point along an MR

boundary for a given point along a histology boundary may be slightly out-of-plane.

Fourth, the in-plane resolution of the registered histological images (0.17 mm pixel width) was approximately four times that of the acquired MR slices (0.70 mm voxel width). This allows partial volume effects that effectively blur several histological pixels into a single MR voxel. In addition, the 3D registration and re-slicing procedure may further reduce the effective in-plane MR resolution through out-of-plane tilting.

However, this effect was minimized by acquiring both histological and MR data approximately perpendicular to the RF electrode path.

Given the above considerations, the small detected differences between corresponding histology and MR boundaries are likely insignificant. Hence, the MR signal probably

accurately identifies regions of tissue destruction.

3.4.3 Comparison of MR Images with Tissue Response

There have been some inconsistent findings with regard to tissue damage at the

margin of MR images of thermal lesions acquired minutes after treatment. Some studies

suggested that a central region of low MR signal corresponds to the region of tissue

damage (10-13). For example, in normal rabbit liver, Lee et al. (10) manually measured

diameters of the central hypointense region in T2-weighted and gadolinium

contrast-enhanced (CE) T1-weighted MR thermal lesions, and measured diameters of a

central region of color change in sectioned fresh tissue. In six lesions, the diameters

agreed within 2 mm. Other studies of ablation showed that a hyperintense rim in MR may

become necrotic (14;25;26). Merkle et al. (14) compared the diameters of a region

occupied by the union of the central zone and surrounding hyperintense zone in T2 and

CE T1-weighted MR lesion images with diameters of coagulation seen in fixed macroscopic tissue slices, without any correction for tissue shrinkage or comparison to histology. These diameters matched within typically 2 mm. From these studies, there is evidence that MR thermal lesion images reflect tissue damage. However, the inconsistencies in the literature indicate the challenge of identifying cell death from color changes in macroscopic tissue sections and from morphological changes in hematoxylin and eosin (H&E) stained histology immediately following ablation. Another potential source of discrepancy is the limitations of diameter measurements for potentially nonsymmetrical lesion boundaries. A careful regional correlation requires the alignment of histology to MR images, and reliable methods to accurately determine the extent of tissue damage. Our studies carefully examined the relationship between cell damage and

MR images on a voxel-by-voxel basis.

A comparison of MR and tissue response a few minutes following RF thermal ablation showed that the central zone M1 and the hyperintense region, zone M2, closely correspond to the region of dead or irreversibly damaged cells in histology. These results are consistent with previous studies that investigated the correlation of MRI and histology following focused ultrasound (25) and laser thermal ablations (26). Using experimental approaches that carefully acquired MR and tissue images in the same plane, and cell-viability staining techniques, these studies suggested that zone M1 corresponds to the necrotic region, and zone M2 will probably become necrotic. However, other investigations that used geometric measurements without alignment of histology to MR images, indicated that only zone M1 correlates with the region of tissue damage (10;13).

These findings indicate that an accurate regional correlation requires reliable methods to determine tissue damage, such as polarized light assessment of muscle damage, and alignment of histology to MR images.

We have determined that the hyperintense rim, zone M2, in the hyperacute CE T1 and

T2-weighted MR images is likely due to increased water from extracellular edema. In histology images, we reliably located a distinct region, zone H2+H3, of increased extracellular space, which strongly indicates edema. Since zone H2+H3 corresponds to zone M2, this is good evidence that edema probably account for the hyperintense rim in the MR images. These results are consistent with previous studies that describe tissue damage following RF ablation (11;27). These studies observed in histology a zone of edema characterized by vacuolation of the neuropil in brain tissue and a zone of small microhemorrhages, but they did not include MR correlation. Our study showed that a region of distinct increased extracellular space, zone H2+H3, that we believe is a region of edematous necrosis, probably gives rise to the hyperintense rim in MR.

It is advantageous to know that the hyperacute MR signal accurately reflects tissue damage and destruction. First, an assessment of the treatment soon after ablation could allow the thermal lesion size and configuration to be adjusted to compensate for deviations from pre-procedural predictions. Second, RF ablation procedures could be extended to the safe destruction of tumor adjacent to vital structures that might be damaged by heating such as the gall bladder, bowel, and especially the brain, where collateral damage must be minimized. Third, MR temperature images could be correlated with MR structural anatomical images, which would assist our development of a mathematical thermal damage model. Fourth, this could eliminate the complicated task of

aligning histology images with in vivo MR images for future studies with tumor. Finally,

since there was no statistically significant difference between hyperintense rim

boundaries for hyperacute T2 and CE T1-weighted MR images, only T2-weighted MR

images could be acquired to reliably monitor the ablation procedure in order to avoid the

cost, potential allergic reactions, and timing issues for uptake and clearance, associated

with a contrast agent, and the time needed for multiple imaging acquisitions.

We conclude that our 3D methodology can be used to accurately map tissue response

to MR thermal lesion images. Observations strongly suggest that in hyperacute T2 and

CE T1-weighted MR images of RF ablated rabbit thigh muscle, the outer boundary of the

hyperintense rim corresponds to the region of eventual cell necrosis within a distance comparable to our ability to measure. This is good evidence that MR thermal lesion images can be used during RF ablation treatments to accurately localize the zone of irreversible tissue damage at the lesion margin.

WORKS CITED

1. Lo, H. W., Tsai, Y. J., Chen, P. H., Chen, H. Y., Ker, C. G., and Juan, C. C. Radiofrequency Ablation for Treatment of Hepatocellular Carcinoma With Cirrhosis. Hepatogastroenterology 2003;50(51):645-50.

2. Ghanem, I., Collet, L. M., Kharrat, K., Samaha, E., Deramon, H., Mertl, P., and Dagher, F. Percutaneous Radiofrequency Coagulation of Osteoid Osteoma in Children and Adolescents. J Pediatr.Orthop.B 2003;12(4):244-52.

3. Livraghi, T., Solbiati, L., Meloni, F., Ierace, T., Goldberg, S. N., and Gazelle, G. S. Percutaneous Radiofrequency Ablation of Liver Metastases in Potential Candidates for Resection: the "Test-of-Time Approach". Cancer 6-15-2003;97(12):3027-35.

4. Farrell, M. A., Charboneau, W. J., DiMarco, D. S., Chow, G. K., Zincke, H., Callstrom, M. R., Lewis, B. D., Lee, R. A., and Reading, C. C. Imaging-Guided Radiofrequency Ablation of Solid Renal Tumors. AJR Am.J Roentgenol. 2003;180(6):1509-13.

5. Mirza, A. N., Fornage, B. D., Sneige, N., Kuerer, H. M., Newman, L. A., Ames, F. C., and Singletary, S. E. Radiofrequency Ablation of Solid Tumors. Cancer J 2001;7(2):95-102.

6. Schenck, J. F., Jolesz, F. A., Roemer, P. B., Cline, H. E., Lorensen, W. E., Kikinis, R., Silverman, S. G., Hardy, C. J., Barber, W. D., Laskaris, E. T., and . Superconducting Open-Configuration MR Imaging System for Image-Guided Therapy. Radiology 1995;195(3):805-14.

7. Cline, H. E., Schenck, J. F., Watkins, R. D., Hynynen, K., and Jolesz, F. A. Magnetic Resonance-Guided Thermal Surgery. Magn Reson.Med. 1993;30(1):98- 106.

8. Gazelle, G. S., Goldberg, S. N., Solbiati, L., and Livraghi, T. Tumor Ablation With Radio-Frequency Energy. Radiology 2000;217(3):633-46.

9. Lewin, J. S., Connell, C. F., Duerk, J. L., Chung, Y. C., Clampitt, M. E., Spisak, J., Gazelle, G. S., and Haaga, J. R. Interactive MRI-Guided Radiofrequency Interstitial Thermal Ablation of Abdominal Tumors: Clinical Trial for Evaluation of Safety and Feasibility. J.Magn Reson.Imaging 1998;8(1):40-7.

10. Lee, J. D., Lee, J. M., Kim, S. W., Kim, C. S., and Mun, W. S. MR Imaging- Histopathologic Correlation of Radiofrequency Thermal Ablation Lesion in a Rabbit Liver Model: Observation During Acute and Chronic Stages. Korean J.Radiol. 2001;2(3):151-8.

11. Merkle, E. M., Shonk, J. R., Zheng, L., Duerk, J. L., and Lewin, J. S. MR Imaging- Guided Radiofrequency Thermal Ablation in the Porcine Brain at 0.2 T. Eur.Radiol. 2001;11(5):884-92.

12. Merkle, E. M., Haaga, J. R., Duerk, J. L., Jacobs, G. H., Brambs, H. J., and Lewin, J. S. MR Imaging-Guided Radio-Frequency Thermal Ablation in the Pancreas in a Porcine Model With a Modified Clinical C-Arm System. Radiology 1999;213(2):461-7.

13. Boaz, T. L., Lewin, J. S., Chung, Y. C., Duerk, J. L., Clampitt, M. E., and Haaga, J. R. MR Monitoring of MR-Guided Radiofrequency Thermal Ablation of Normal Liver in an Animal Model. J.Magn Reson.Imaging 1998;8(1):64-9.

14. Merkle, E. M., Boll, D. T., Boaz, T., Duerk, J. L., Chung, Y. C., Jacobs, G. H., Varnes, M. E., and Lewin, J. S. MRI-Guided Radiofrequency Thermal Ablation of Implanted VX2 Liver Tumors in a Rabbit Model: Demonstration of Feasibility at 0.2 T. Magn Reson.Med. 1999;42(1):141-9.

15. Thomsen, S., Pearce, J. A., and Cheong, W. F. Changes in Birefringence As Markers of Thermal Damage in Tissues. IEEE Trans.Biomed.Eng 1989;36(12):1174-9.

16. Thomsen, S. Pathologic Analysis of Photothermal and Photomechanical Effects of Laser-Tissue Interactions. Photochem.Photobiol. 1991;53(6):825-35.

17. Breen, M. S., Lancaster, T. L., Lazebnik, R. S., Nour, S. G., Lewin, J. S., and Wilson, D. L. Three-Dimensional Method for Comparing in Vivo Interventional MR Images of Thermally Ablated Tissue With Tissue Response. J Magn Reson.Imaging 2003;18(1):90-102.

18. Lazebnik, R. S., Lancaster, T. L., Breen, M. S., Lewin, J. S., and Wilson, D. L. Volume Registration Using Needle Paths and Point Landmarks for Evaluation of Interventional MRI Treatments. IEEE Trans.Med.Imaging 2003;22(5):653-60.

19. Lancaster, T. L and Wilson, D. L. Correcting Spatial Distortion of Histological Images. Annals of Biomedical Engineering 2002;(In Press).

20. Chalana, V. and Kim, Y. A Methodology for Evaluation of Boundary Detection Algorithms on Medical Images. IEEE Transactions on Medical Imaging 2002;16(5):642-52.

21. Nour, S. G., Aschoff, A. J., Mitchell, I. C., Emancipator, S. N., Duerk, J. L., and Lewin, J. S. MR Imaging-Guided Radio-Frequency Thermal Ablation of the Lumbar Vertebrae in Porcine Models. Radiology 2002;224(2):452-62.

22. Aschoff, A. J., Rafie, N., Jesberger, J. A., Duerk, J. L., and Lewin, J. S. Thermal Lesion Conspicuity Following Interstitial Radiofrequency Thermal Tumor Ablation in Humans: a Comparison of STIR, Turbo Spin-Echo T2-Weighted, and Contrast- Enhanced T1-Weighted MR Images at 0.2 T. J.Magn Reson.Imaging 2000;12(4):584-9.

23. Graham, S. J., Stanisz, G. J., Kecojevic, A., Bronskill, M. J., and Henkelman, R. M. Analysis of Changes in MR Properties of Tissues After Heat Treatment. Magn Reson.Med. 1999;42(6):1061-71.

24. Cutts, A. Shrinkage of Muscle Fibres During the Fixation of Cadaveric Tissue. J.Anat. 1988;160:75-8.

25. Chen, L., Bouley, D. M., Harris, B. T., and Butts, K. MRI Study of Immediate Cell Viability in Focused Ultrasound Lesions in the Rabbit Brain. J.Magn Reson.Imaging 2001;13(1):23-30.

26. Morrison, P. R., Jolesz, F. A., Charous, D., Mulkern, R. V., Hushek, S. G., Margolis, R., and Fried, M. P. MRI of Laser-Induced Interstitial Thermal Injury in an in Vivo Animal Liver Model With Histologic Correlation. J.Magn Reson.Imaging 1998;8(1):57-63.

27. Farahani, K., Mischel, P. S., Black, K. L., De Salles, A. A., Anzai, Y., and Lufkin, R. B. Hyperacute Thermal Lesions: MR Imaging Evaluation of Development in the Brain. Radiology 1995;196(2):517-20.

CHAPTER 4

Registration of Subacute MR Lesion Images to Histological Sections with Model-Based Evaluation

4.1 INTRODUCTION

There are several reports on experimental methods to correlate medical images with

tissue features. As an example application is the comparison of subacute (four days

post-ablation) magnetic resonance (MR) thermal lesion images with the tissue response

(1-6). Only one report used registration, however it was a two-dimensional (2D) method

(6). Others simply compared thermal lesion diameters measured in 2D MR images with

diameters measured in macroscopic tissue images (1;3-5) or histology images (2). Such

methods are limited by ones ability to accurately determine corresponding image slices

and by the applicability of simple length measurements to capture what might be a

complicated three-dimensional (3D) geometry.

A 3D method for aligning histology and medical images avoids the problems

identified above. Using a 3D volume registration technique, one can obtain a voxel to voxel correlation throughout the volume of interest. For our application to evaluate a minimally invasive interventional MRI (iMRI) thermal ablation procedure, we can address the practical clinical problem of determining if MR images of thermal lesion correlate with tissue response in tissue containing tumor. For separate geometric measurements in MR images and sliced tissue, slicing a specimen in the proper plane is difficult and measurements are limited to symmetrical lesion boundaries. Alignment of

histology and MR images is necessary for tumor tissue with irregular boundaries. For 2D alignment methods, orienting and slicing an organ in the correct plane is problematical, and a technique for correcting misaligned slices is unclear.

Though there are several reports on 3D volume registration, most are inappropriate for our applications. MR images and color photographs of histological sections from excised tissue are very dissimilar, especially those from muscle tissue, as shown later.

Grayscale registration methods, such as mutual information maximization, rely on the presence of at least some common features in both volumes (7). Surface matching techniques are inappropriate due to potential warping of the external surface or potential loss of significant regions during tissue excision. Furthermore, surface landmark based registration is inappropriate for our application because the skin is commonly removed to aid tissue fixation. Point landmark based methods, utilizing internal anatomical or artificially implanted features, are feasible but require the localization of landmarks common to both volumes. Internal point landmarks are often difficult to locate accurately in both volumes, due to voxel dimensions and/or contrast limitations. Registration based on extended anatomical features such as lines and planes (8) is not feasible in our experiments because it is difficult to identify suitable landmarks in quite homogeneous muscle tissue.

Since there are special requirements in our application, we desired a more flexible approach suitable for a wide variety of tissues, spatial distortions, and missing image data due to tissue tearing. Some methods use relatively few control points, and we require many, especially at edges, in order to correct spatial distortions found there. This is particularly an issue because thermal ablations are often applied near the surface of an

organ. Jacobs et al. used a surface-based, “head in hat,” 3D method to register rat brain histology and MR images (9). To correct for spatial distortion in the histological images, they warped 2D MR images to the histological samples. Radial correspondence from a center of mass was assumed to obtain corresponding points for the thin plate spline (TPS) transformation. Although radial correspondence is reasonable to apply in the brain, it is not generally applicable in other types of tissue. Grayscale matching methods require at least somewhat similar features, and the brain data used by many has an abundance of such features (10). Macroscopic tissue images, especially those of rabbit thigh muscle, have very few features suitable for matching to histology. Finally, our histology samples are often missing sections due to tearing, probably because ablations can change the consistency of the tissue sample and tissue such as muscle can contain anatomical boundaries that separate during processing. For these reasons, we desired a 2D method which can be used interactively in a wide variety of preparations.

In this paper, we describe our method for the direct comparison of subacute MR images to histology. Our method consists of a 3D rigid body registration of a MR volume to macroscopic tissue photographs using needle fiducials followed by a 2D warp of histological sections to the macroscopic tissue photographs to correct deformation that occurs during tissue processing. Our 3D method requires both computer registration methods and tissue handling techniques. We apply the method in radio-frequency (RF) thermal ablation experiments under iMRI guidance in an animal model of rabbit thigh muscle. We present several validation results to show the accuracy of the method. A comparison of MR and tissue response is then described.

4.2 MATERIALS AND METHODS

4.2.1 Registration Procedure

To correlate histological and 3D MR image data, we use semi-automatic registration

algorithms as well as special tissue handling procedures. Here we briefly review the

entire approach so as to introduce the details of the computer registration methods.

Experimental details are described in subsequent sections.

T2-weighted MR image volumes are obtained. At least two fiducial needles are

included near the region of interest that can be seen in the MR images and leave a hole

that can be found in the tissue. An approximately 45° angle is used between the fiducial

needles to support a 3D registration. After fixation, we slice and photograph the tissue block on a special custom-made apparatus at 3 mm intervals to obtain a volume of reference tissue images. Histology samples are obtained in a plane parallel to these tissue images and digitized on a video microscopy system. Using two different computer algorithms, we register the MR and histology images to the reference tissue images to facilitate a voxel to voxel comparison between medical and histology images.

4.2.2 3D Registration of MR Images to Reference Tissue Images

We consider two volumes of images: a reference volume, which remains unchanged, and a non-reference volume, which will be re-sliced using volume interpolation. Both volumes feature the same fiducial needles and optional fiducial points. Registration consists of manual selection of points and needle paths and automatic matching of these features using an iterative optimization technique.

Manual identification of features is performed using a preprocessing program having

a graphical user interface. The program allows an operator to view cross sections of the

unregistered volumes side by side and choose the point in each slice where the needle

intersects. Each identified point is assigned a 3D spatial location, in millimeters, relative

to an origin located in the upper left corner of the volume’s first slice. For photograph

volumes, the z coordinate is assumed to be that of the exposed face of each slab, but for

MR volumes, it is assumed to lie midway through the slice thickness.

After manual identification is complete, we initiate an automated optimization of the

registration parameters required for 3D rigid body transformation (translations and

rotations about X, Y, and Z) and uniform scale. For this we apply the iterative closest

point algorithm, where the reference and non-reference volumes are considered the model

and data respectively (11). First, we create a parametric representation of each needle

path in the non-reference volume by projecting each needle onto the XZ and YZ planes

and fitting a line to the data in each plane using a least-squares criterion.

We next determine corresponding point pairs between the unregistered volumes. For

each of the h point landmarks, the corresponding point is chosen by the operator. For q

needle pairs, we examine each chosen point p along a needle in the reference volume and

assume that the closest point p’’ along the non-reference fitted line corresponds. We

compute a vector ĝ from any point p’ along the line to p, and a unit vector û along the

line. The closest point is p’’= p’ + (ĝ • û) û. This equation is exact for straight lines; an iterative solution is required for more complex curves. As registration improves with iterative optimization, so does the validity of the assumption that the closest point is the true corresponding point (11).

We now define a scalar objective function that represents registration error in terms of the distances dp between each of the h point landmark pairs and dn between each of the

ti points along each of the q needles.

q t h i n 2 p 2 f = ∑∑(di, j ) + ∑(dk ) (4.1) i==11j k =1

We utilize the closed form quaternion-based approach described by Besl and McKay

(11) to compute the rigid body transformation matrix that will minimize f for the current

set of correspondence points. We then apply the matrix to the non-reference volume point

landmarks and needle points and repeat the above procedure. Because we take care to

acquire images in a similar geometric fashion, the initial guesses for all transformation

parameters relative to the reference are all zeros. A tolerance of 1E-8 is the termination

criterion. The optimized transformation matrix is used to reslice the non-reference

volume using tri-linear interpolation. For registration to a photograph tissue volume, the

photographs are the reference and image slices from a medical scanner are interpolated.

The method was previously validated by Lazebnik et al. (12) using computer

phantoms and brain tissue images, and errors were estimated to be less than 0.9 mm.

4.2.3 2D Registration of Histology to Reference Tissue Images

We cut each thick-sectioned tissue slice to fit a 4.0 x 4.0 cm histology-mounting block.

Images of these “tissue sub-sections” are obtained using a digital camera and copy stand,

and aligned to their corresponding reference tissue images using a 2D affine

transformation. In each image, we use a cursor to select at least three correspondence

points such as the fiducial needle tracks, intersections of fat boundaries, and corners. We

calculate the least-squares solution and transform the tissue sub-section image to match the tissue image.

To align the histology image with this transformed tissue sub-section image, we developed a 2D TPS warp algorithm (13;14). Using a cursor, we choose correspondence points by selecting the interior ink marks and anatomical landmarks such as blood vessels, and intersections of muscle group boundaries. We also obtain corresponding points along external boundary segments. Corresponding boundary segments are obtained by selecting readily identifiable end points consisting of fiducial ink marks and corners.

A computer algorithm obtains a curve representing the border segment, measures the distance between the end points, and creates typically 5-10 equidistant points along the curve in both images. These provide additional correspondence points to match. Often multiple border segments are used. All correspondence points are included to compute the TPS transformation.

The segmentation of border segments is accomplished using a modified live-wire algorithm (15). To apply live-wire segmentation, the user inputs a starting point and, as the cursor is moved along the boundary, the path snaps to the edge like a “live-wire” until a final end point is chosen and segmentation is completed. This is a natural choice for our interactive method requiring segmentation of identifiable boundary segments. Inputs to the algorithm consist of a starting point and a gradient magnitude based cost map of the image. Cost maps for histological images are created by applying a Sobel filter to the median filtered green channel. A 3x3 median filter is applied to fill small spaces that naturally occur in histology samples while preserving boundary edges. For the tissue image, a Sobel filter is applied to the red, green, and blue channels, and the maximum

gradient magnitude of the three is used as the cost at each pixel in the image. Cost maps

are calculated once and stored to speed the interactive portion of the algorithm. A four-

connected, 2D dynamic programming algorithm is used to determine the minimum cost

and its path from the starting point to any possible end point contained within a region of

interest in the image. Once this is completed, as the user interactively moves the cursor to

potential end pixels, the program can quickly backtrack to the starting point giving the

impression of a live-wire. If the path strays from the desired boundary, the additional low

cost control points can be added to anchor the path to the boundary.

Once correspondence points are established, a TPS transformation is applied to align

the histology image to the gross tissue photograph. For the TPS transformation we obtain

a mapping

′ ′ g[x, y] = f [x , y ] = f [a[x, y],b[x, y]] (4.2) where g[x,y] represents the pixel in the output image at coordinate (x,y) and f[x’,y’] is the value at (x’,y’), a real-valued location in the input image. To perform the TPS transformation, we determine parameters in the functions a[x,y] and b[x,y] from correspondence points. Then, for each (x,y) location in the output image, we compute

(x’,y’) in the input image. To obtain appropriate values at real-valued locations in the input image, we use bilinear interpolation.

The TPS algorithm is described by Bookstein (13) and implemented using the algorithm described by Davis et al. (14). The TPS warp is based on the deformations of a

thin metal plate, and provides a smooth, continuous transformation. TPS does not involve

a least-squares solution; instead it is an interpolating function, which guarantees that

correspondence points match exactly.

This method was previously validated by Lancaster et al. (16) using extra ink fiducials in tissue images, and mean errors were estimated to be 0.6 mm for brain tissue, which works better than some other tissues.

4.2.4 Experimental Methods

The correlation of in vivo MR and histological image data require careful animal and tissue handling methods. The experimental methods include tissue ablation, MR imaging, tissue slicing and photographing, and histological processing; all were very important for successful registration. It is extremely important to minimize tissue deformation and destruction during the dissection and slicing processes for accurate registration. The details and the validation experiment for the registration are described in the next sections.

4.2.5 RF Ablation in Rabbit Model and MR Imaging

Following a protocol approved by the Institutional Animal Care and Use Committee, we anesthetized four New Zealand White rabbits. We shaved the left thigh of each rabbit and placed them within the gantry of a clinical 0.2 T C-arm MR imaging system (Siemens

MAGNETOM OPEN, Erlangen, Germany). We secured the legs of each rabbit to a customized Plexiglas support to prevent thigh movement. After positioning two ground pads on each rabbit’s abdomen, we placed the thighs within a 12 cm diameter multi-turn solenoid, receive-only coil.

We performed the RF ablation procedure under MR guidance. We inserted an

MR-compatible RF electrode with a 10 mm exposed tip percutaneously into the thigh

muscle. Lesion formation was achieved by increasing the local temperature by delivering

RF electric current between the electrode tip and ground pads. We applied RF energy for

2 minutes using a 100 W generator operating at 500 kHz. The tip of the RF electrode was

maintained at a temperature of 90 ± 2°C using a thermistor within the electrode tip to

provide accurate instantaneous temperature information. Four days later, the animal was

re-anesthetized and at least two MR-compatible 22 G, 10 cm fiducial needles were

inserted in the lesion vicinity with one fiducial approximately parallel and the other

fiducial approximately at an angle of 45° to the RF electrode track. A T2-weighted

turbo-spin-echo sequence was applied with TR/TE/NSA parameters of 3362/68/8 that

gives 256 x 256 x 9 voxels over a 180 x 180 x 27-mm FOV to yield 0.70 x 0.70 x 3.0-mm

voxels oriented to give the highest resolution for slices approximately perpendicular to

the RF electrode. This imaging plane minimizes partial volume error because temperature

changes little along the needle electrode. The rabbit was sacrificed post imaging using a barbiturate overdose technique via intravenous administration of 0.1 ml/lb of pentobarbital sodium. After the tissue was fixed in formalin, we examined the tissue features in histological sections.

4.2.6 Tissue Slicing and Acquisition of Calibrated Tissue Images

A tissue slicing apparatus was constructed that included a tissue platform and a digital

camera (DSC-D770, Sony, Japan). To prevent any foreshortening in the tissue

photographs, we ensured that the imaging plane of the camera was parallel to the tissue

face. We sliced the tissue in planes approximately perpendicular to the fiducial needle

orientated approximately parallel to the RF electrode. To orient the specimen and reduce

deformation during slicing, the specimen was positioned in a Styrofoam block and held in place with tissue embedding wax (Parablast X-tra, Oxford Labware, St. Louis, MO). A black coloring agent was added to the translucent wax to create a clear tissue border in the photographs. The Styrofoam block was secured to the tissue-slicing platform with glue and tape.

We obtained 3 mm tissue slices by using the specially designed apparatus that included a linear displacement device (Rack and Pinion Slide, Edmund Scientific,

Barrington, NJ) for accurate stepping of the platform in small increments. We sliced the specimen with a 12.8 inch autopsy knife (Tissue-Tek Accu-Edge Semi-Disposable

Autopsy Knife System, Sakura Finetek, Japan) using the vertical supports of the slicing apparatus as a guide. Repeatedly, we photographed the tissue block face, advanced the platform by 3 mm, and sliced the tissue, until the specimen was traversed. The tips of fiducial needles were exposed at the tissue block face, highlighted with ink for easy identification in each photograph, and stepped back slightly beyond the plane of the next tissue slice. We calibrated the scale of macroscopic tissue images using a ruler in the plane of the tissue slice. To maintain calibration, the imaging geometry and the manual zoom on the digital camera were fixed during acquisitions.

For the last step, we prepared the tissue slices for histological processing. We cut each thick-sectioned tissue slice to fit a histology-mounting block. After placing ink-mark fiducials at the edge and interior of these tissue sub-sections with pins dipped in colored ink (Davidson Marking System, Bradley Products, Bloomington, MN), we photographed them on a copy stand. The color-coded head of each pin gave a 2 mm diameter disk in the image that enabled us to accurately localize the center of the pin. We embedded the tissue

slices in paraffin and obtained Masson trichrome (MT) stained histological sections near the photographed tissue face and parallel to the tissue images.

4.2.7 Acquisition of Histology Images

We digitized the histology slides using a video microscopy system with a motorized stage. The system consisted of a light microscope (BX60, Olympus, Japan), video camera

(DXC-390, Sony, Japan), position encoded motorized stage (ProScan, Prior Scientific,

Rockland, MA), and software (Image-Pro with Scope-Pro, Media Cybernetics, Silver

Spring, MD). To obtain an image of the entire slide, we used the tiling function of

Scope-Pro. This procedure drove the motorized stage, acquired a series of pictures, and seamlessly combined the photographs to form one large tiled image. We performed tiling with a 4x objective. Typically, the tiled image consisted of about 100 image acquisitions over a 1.0-1.5 cm lesion and gave about 6400 x 4800 square pixels that were 5.21 um on a side. By reducing the tiled image to 10 percent of its original size, we created a smaller map image on which we marked the locations of ink fiducials. Using the motorized stage under joystick control, we determined the position of each fiducial on the slide with a digitally imposed crosshair at the center of the live video window. The stage coordinates were used to automatically mark the correct fiducial positions on the map image with colored graphical overlays.

4.2.8 Validation Methods

We visually evaluated alignment of internal structures, such as boundaries of fat and the thermal lesion, in MR, tissue, and histology images, using Regviz, a program written in

interactive data language (IDL) (Research System Inc., Boulder, CO) and created in our laboratory for visualizing and analyzing registered image volumes. This program has a linked cursor that allows us to compare locations within multiple registered images.

Using this program, we outlined various anatomical features in reference tissue images and copied them to corresponding images from histology and MR. It also allowed sectored displays where sections from different images were tiled together to show continuation of boundaries. Finally, the program allowed an overlay of a grayscale image from one type of acquisition to a “hot iron” color map from another. These features allowed us to visually evaluate registration over aligned volumes of images.

We evaluated the registration accuracy using extra fiducial needle paths not optimized during alignment. For every registered MR and histology image, we manually localized the center of the extra needle paths. For the MR images, we assumed needles were straight and obtained a least-squares fitted line from these points. For each point from histology images, we established a corresponding point for the fitted line in MR based on the closest 3D point along the fitted line. A 3D alignment error was determined by calculating the mean distance between corresponding points along the needle.

4.2.9 3D Model-Based Validation

Since preliminary analyzes indicate that surfaces visible in subacute MR lesion and histology images are very probably identical, we will test potential errors in registration by determining the distances between these surfaces. We applied a geometric 3D deformable lesion model to registered in vivo MR and histology image volumes. For the

MR image volume, we used a 12-parameter model of two concentric ellipsoids that are

simultaneously fit to the inner and outer surfaces of the lesion in the in vivo MR image volumes. For histology, we used a 9-parameter model with one ellipsoid fit to the surface of a distinct necrotic region surrounding the lesion area. The model is fitted to manually segmented lesion images through iterative minimization of an objective function that describes the average distance between model surfaces and corresponding segmented points. The model parameters and fitting procedure are described elsewhere (17). The method was previously validated by Lazebnik et al. (17) using computer phantoms and in vivo lesion images, and median model fit errors were estimated to be no more than

0.58 mm.

We determined the 3D registration error by calculating the distance between the MR and histology 3D model surfaces using a software program written in

MATLAB (Mathworks, Natick, MA). An automatic algorithm determined equally spaced points, 0.7 mm apart, along the histology ellipsoid model surface. For each such reference point, we computed the 3D Euclidean distance to closest point on each MR ellipsoid model surface. If the reference point was inside a MR model surface, the distance was negative, else it was positive. This allowed us to determine if the histology surface was interior or exterior to MR. We also measured the displacement of the ellipsoid centroids to estimate the 3D registration error.

To fit the ellipsoid models, we manually segmented MR and histology lesion images.

An observer segmented boundaries in registered histological slides under the supervision of a pathologist. For these experiments, we marked the boundary in MT histology samples that sharply divided a central necrotic zone, hereafter referred to as H1, with coagulated and blue stained cells from surrounding normal muscle tissue with red stained

cells. This boundary was marked in every histology image using the method previously

described to mark ink fiducials with the video microscopy system. The observer

segmented histology boundaries without knowledge of MR boundaries. For each

boundary, we marked 15-30 closely spaced points on the histology map image with

colored graphical overlays, yielding a boundary that can be compared with a boundary

marked in registered MR images. These map images were registered to their

corresponding tissue images.

In the registered T2-weighted MR lesion images, we manually segmented the central

isointense/hypointense region and the outer boundary of the surrounding hyperintense

rim, hereafter referred to as M1 and M2, respectively. An observer independently performed the segmentation using a freehand region of interest (ROI) tool in Analyze

(Analyze Direct, Lenexa, KS) image analysis software. This observer segmented MR images without knowledge of histology boundaries. Important software features included the ability to look at adjacent images slices simultaneously and a ROI editing tool. The observer followed a strict protocol. All segmentation was performed on the same workstation in a darkened room. We perceptually linearized the display using Optical

Following segmentation, the boundary coordinates were exported from Analyze and examined using MATLAB. Each segmented point was assigned a 3D coordinate, in

millimeter dimensions, based on the voxel dimensions. The set of boundaries was utilized for subsequent model fitting and error estimation.

4.3 RESULTS

Figure 4.1 shows a typical in vivo T2-weighted MR lesion image well registered with macroscopic tissue and histology images. We copied the H1 boundary to the registered tissue and MR images. The H1 region was characterized by muscle cells with loss of nuclei and contraction band necrosis (focal band-like coagulation of muscle contractile elements). This boundary matched features well in the tissue and MR images. The outer boundary of the dark brown rim in the tissue image corresponded well to the H1 boundary. In the MR image, the H1 and M2 boundaries were closely aligned. Graphical overlays placed around needle tracks in the MR images were copied to tissue images where they were aligned with needle tracks in tissue and histology images. For two specimens which had extra fiducial needles to assess accuracy, the mean 3D distance between the needle paths was 1.32 ± 0.39 mm (mean ± SD), which compared favorably to the MR voxel dimensions (0.70 mm in-plane and 3.0 mm thick). We obtained comparable results with needle paths optimized during registration, with a distance error of 1.27 ± 0.36 mm. For a volume registration with eight histological samples, scatter plots of the x, y, and z displacements, as function of tissue slice, appeared random. The mean for each displacement was always less than 0.18 mm, a value close to zero. This is good evidence of random error rather than a systematic misregisration.

Figure 4.1. Registration quality for MR, tissue, and histology images. Images are: in vivo T2-weighted MR (a,b), macroscopic tissue (c,d), and MT histology (e,f). The sam e images are shown in the left and right columns. The histological sample is smaller since we cut each thick-sectioned tissue slice to fit the histology-mounting block. The therm al lesion is the elliptical region that has an isointense/hypointense central core (M1) surrounded by a hyperintense rim (M2) in MR, a dark brown rim in tissue, and a purple stained region of necrosis (H1) in histology. The H1 boundary (yellow) was marked and copied to macroscopic tissue and MR images, where it matched features in these images. Two needle tracks used for registration (circles) and two used for validation (squares) were marked in the MR image and copied to tissue and histology images. The needle tracks are dark regions surrounded by three bright spots, a characteristic artifact, in MR images, small dark regions in tissue photographs (arrows), and white holes in histology images (arrows). Occasionally, a needle track will be positioned in unstained fat tissue between different muscle groups and not be visible in a histology image (square in bottom right image without arrow). Excellent registration accuracy of MR, tissue, and histology images is clearly evident with good correspondence of needle tracks. In addition, a muscle group separation of fat, which is a bright curve in MR and an unstained curve in histology, was marked in MR (green arrows) and copied to histology image, where it matched features. Results were similar in several adjacent tissue slices.

To further analyze the correspondence between histology and MR lesion boundaries, we manually segmented M1 and M2 boundaries and compared them with the H1 boundary. In Figure 4.2, the H1 and M2 boundaries were well aligned, which indicated excellent registration quality of the MR and histology images. Results were remarkably similar across several adjoining tissue slices.

Figure 4.2. Comparison of H1 boundary with M1 and M2 boundaries. The registered images are: in vivo T2-weighted MR (a,b) and MT histology images (c,d). The images in the left and right column are identical except for graphical overlays on the right showing the

M1 (green), M2 (red), and H1 (yellow) boundaries. The H1 boundary was marked on the histology image and copied to MR image, where it closely matches the M2 boundary. A histology image with only a partial boundary was shown because of the excellent contrast of the stain in histology.

We qualitatively evaluated the fit of the ellipsoid model to histology and MR image data. In Figure 4.3, we compared the fit of the model surface to the manually segmented

M1, M2, and H1 boundaries visualized in three-dimensions. We displayed the boundary points and compared them with a surface rendering of the corresponding model surfaces.

These typical ellipsoid model fits demonstrated good correspondence between the model surfaces and segmented boundaries.

Figure 4.3. Fit of the ellipsoidal model to segmented lesion boundaries, and visualized in three dimensions (units indicate millimeters). Regularly spaced points along segmented boundaries, for every applicable section, define the M1 and M2 boundaries (a), and the H1 boundary(c). The missing histological boundary data was due to muscle’s anatomical boundaries that occasionally separate during processing. Af ter the appropriate model is fit, the two concentric ellipsoid model in MR (b) and one ellipsoid model in histology are visualized with a surface rendering. For the two ellipsoid model, we display a cut-away of the outer surface to reveal the inner ellipsoid. The continuous model surfaces closely approximate the actual lesion geom etry as defined by the segmented data.

We also quantitatively analyzed the fit of the model to histology and MR image data.

We examined the signed distance between segmented boundary points and the corresponding model surface where a point inside or outside an ellipsoid model volume has a negative or positive sign, respectively. Collapsing over all specimens for the M1,

M2, and H1 boundaries respectively, the mean signed distance was 0.00 ± 0.00 mm

(mean ± SD), 0.00 ± 0.00 mm, and 0.03 ± 0.05 mm, indicating insignificant bias between corresponding segmented boundaries and model surfaces. In Figure 4.4, we plotted the mean absolute distance as a function of the specimen. For the M1, M2, and H1 boundaries respectively, the mean absolute distance was 0.54 ± 0.09 mm,

0.70 ± 0.07 mm, and 0.50 ± 0.11 mm, which compares favorably to the in-plane MR voxel dimension (0.70 mm) and slice thickness (3.00 mm). Overall, the modeling error is small.

Figure 4.4. Plotted as a function of specimen are absolute distances from manually segm ented lesion boundaries to the fitted model surface. Mean and standard deviations values for M1, M2, and H1 boundaries are shown. Each data point represents an average over typically 100 equally spaced points along a boundary and usually five boundaries for each specimen.

We evaluated the correspondence of the ellipsoid model from registered MR and

histology images. In Figure 4.5, we compared the M2 and H1 model surfaces visualized

in three-dimensions. We combined the surface renderings of both model surfaces and

varied the transparency to visualize the internal surface. The accurate 3D registration of

MR to histology images is evident by the close alignment of these MR and histology

model surfaces.

Figure 4.5. Fusion of the MR and histology ellipsoidal model surfaces, and visualized in three dimensions (units indicate millimeters). After the appropriate model is fit to registered MR and histology segmented boundaries, the M2 (red) and H1 (blue) ellipsoid models are visualized with a surface rendering. Since the M2 and H1 model surfaces are 50 percent transparent and opaque, respectively, we can examine the regions of the H1 model surface that are inside the region bounded by the M2 ellipsoid volume. An H1 model surface inside the M2 model surface is purple, else it is blue. Accurate registration accuracy of MR and histology images is evident with good correspondence of the continuous M2 and H1 model surfaces.

Using models fitted to the registered image data, we determined the discrepancy

between the M2 and H1 model surfaces. In Figure 4.6, we plotted the signed and absolute

distance for model surfaces as a function of the specimens. Collapsing over all

specimens, the mean signed distance was -0.06 ± 0.19 mm, a value close to zero,

indicating insufficient bias between model surfaces. The mean absolute distance of model

surfaces and centroids over all specimens was 0.96 ± 0.13 mm and 1.19 ± 0.31 mm,

respectively. This small discrepancy compares favorably to the MR voxel dimensions

(0.70 mm in-plane and 3.00 mm thick).

Figure 4.6. Plot as a function of specimens are the mean signed (a) and absolute (b) distances between M2 and H1 model surfaces. The average and standard deviation are calculated over typically 1300 equally spaced points, 0.7 mm apart, along the H1 model surface. Signed distan ce is negative when the H1 surface is inside the M2 ellipsoid volume, else it is positive.

In Figure 4.7, a careful examination of the relative shape, and orientation of the

closely aligned H1 and M2 boundaries indicates that these boundaries may align perfectly

with a small rotation and translation of the MR volume. To examine this, we registered

the surfaces by setting the values of centroid and three rotation model parameters in the

M2 ellipsoid equal to the values in the corresponding H1 model. In Figure 4.8, we plotted

the signed and absolute distance for the registered model surfaces as a function of the

specimens. Collapsing over all specimens, the mean signed distance was 0.24 ± 0.19 mm, a value close to zero. The mean absolute distance between the model surfaces over all specimens was 0.37 ± 0.12 mm, a decrease of approximately one-half. Hence, the M2 and H1 segmented boundaries probably achieve sub-voxel accuracy.

Figure 4.7. Correspondence of segmented H1 and M2 boundaries. The T2-weighted MR im ages (a,b) are identical except for graphical overlays on the bottom showing the H1 (yellow) and M2 (red) segmented boundaries. The H1 boundary was marked on the histology image and copied to the registered MR image. Note the remarkably similar shapes of the closely corresponding H1 and M2 boundaries. Results were similar in several adjacent tissue slices.

Figure 4.8. Plotted as a function of specimen are mean signed (a) and absolute distances (b) between registered H1 and M2 model surfaces. The average and standard deviation are calculated over typically 1300 equidistant points along the H1 model surface. Signed distance is negative when the H1 model surface is inside the M2 ellipsoid volume, else distance is positive.

4.4 DISCUSSION

4.4.1 Registration Assessment Methodology

Relatively little validation work has been done to show that medical and histological images are correctly aligned. Visual methods such as contour overlays and difference images are most often employed (18-21). In those cases where point anatomical

landmarks are visible in both image pairs, one could estimate error, but this has not been widely reported.

The distance between the ellipsoid model surfaces and volume centers to assess the error of the 3D registration is sound for several reasons. First, the M2 and H1 model surfaces are most likely identical since the mean absolute distance between surfaces is

0.37 ± 0.12 mm after registering the model surfaces, which compares favorably to the in- plane MR voxel dimension (0.70 mm) and thickness (3.0 mm). Hence, the surfaces achieve sub-voxel accuracy. Second, the 3D surface model provides a good approximation to the actual lesion geometry. In a previous study, we developed and validated an ellipsoid model that described the hyperintense region of in vivo MR lesions and quantitatively determined that this model provided a good fit to manually segmented boundaries (17). This technique was shown to be resistant to image noise and missing segmentation information while allowing the resulting extracted boundaries to achieve sub-voxel accuracy. For all in vivo lesions, the median distance from the model surface to segmented boundaries was no more than 0.58 mm, less than a voxel width of 0.7 mm and much less than the slice thickness of 3.0 mm. Third, this independent measure of registration quality is attractive since it can be applied to any thermal lesion image data, even those that do not have extra fiducial needles.

Our method for estimating the 3D registration error with extra fiducial needles can be practically applied. We can reliably localize the fiducial needle tracks in histology, tissue, and in vivo medical images such as MR and CT data, even those that do not have thermal lesions. There are no 3D anatomical point landmarks and very few anatomical landmarks in our most common preparation, rabbit thigh muscle. Surface point fiducials attached to

the thigh muscle will not be found in histological samples. We have not determined how to add internal point fiducials that can be reliably localized in tissue and histological specimens, and not dislodged during tissue slicing or missed between thick tissue slices.

By adding additional needles not used in the registration, an independent measure of registration quality is obtained. The biggest problem with using needle track displacement as a measure of error is that it also includes needle localization uncertainty.

For histology and tissue images, the uncertainty was probably negligible. However, for

MR images, Lewin et al. (22) showed that the error in localizing a needle with our MR system and scanning sequence was typically within 1.0 mm. Hence, as much as 1.0 mm of our 1.3 mm average registration error based on displacement of extra needles might be due to needle track localization error.

4.4.2 Assessment of Fiducial Localization Error

The localization error of ink fiducials in the macroscopic tissue images was minimized by using the centroid calculated from the pinheads tissue sections. In histology images, one cannot always identify the exact center of ink marks, which often show up as slits in the tissue lined with colored ink. Maximum slit size can be on the order of 1 mm, and by choosing the center of the slit, we have a maximum possible error of 500 µm. Using anatomical landmarks could reduce this error, especially in a sample rich with landmarks such as the brain.

In order to minimize needle localization uncertainty for MR images, users of our method must familiarize themselves with the MR appearance of fiducial needles. The magnetic properties of a metallic needle result in a susceptibility artifact that can distort

its shape. Consequently, the cross sectional appearance of a fiducial needle oriented

approximately perpendicular to the image plane is a hypointense central region and three

hyperintense points in approximately triangular configuration about the center, while

needles oriented more parallel to the image plane result in a hypointense streak bordered by a hyperintense rim. For our interventional MRI systems, the actual needle center is very well approximated by the center of the hypointense region of the artifact (23). For our in vivo imaging parameters, this area was approximately three voxels across, and thus we expect the error to be at most a voxel width. Finally, as our registration results indicate, high quality registration is feasible despite this consideration.

4.4.3 Limitations

The registration method work reliably after several challenges, such as warping of the tissue due to movement and tearing of the tissue during dissection and slicing, were overcome with improved slicing equipment and experience. The 3D registration method also compensated for distortions from slight tissue shrinkage due to fixation by using a uniform scale parameter in the needle registration. For these thigh muscle experiments, shrinkage was typically five percent, a value consistent with previous studies (24), which showed a mean shrinkage of 2.3 percent for muscle tissue. In addition, a potential source of local distortion might occur with any tissue swelling due to edema. However, such swelling should occur before insertion of the needle fiducials, and any regional distortion should be consistent across the image data.

It is understood that if the needles are placed parallel to one another, there is no constraint for registration along the z-axis. Previously, we performed many simulation

experiments and determined that the method was quite accurate as long as the angle between the needles exceeded 15º, even with a needle localization error (12). We use an approximately 45º angle.

4.4.4 Correlation of Subacute MR Lesion Images with Tissue Response

We have shown that the M2 boundary closely corresponds to the H1 boundary of tissue destruction four days post-ablation. These results are consistent with previous studies that investigated the correlation of MRI and histology immediately following radio-frequency

(5), focused ultrasound (2) and laser ablations (6). Based on 2D registration and diameter measurements, these studies suggest that M1 region corresponds to a necrotic region, and the M2 will possibly become necrotic. Since we showed in a previous study that the M2 boundary changed insignificantly after four days (25), the M2 boundary is a reliable marker of cell death.

We conclude that our methodology can be used for accurate tissue typing to relate 3D medical images to tissue features. Using the method to evaluate RF thermal ablation, we found good spatial correspondence between registered in vivo MR lesion images and histology samples. This is good evidence that using this method to incorporate histological data with MRI can be an important tool for evaluating and improving interventional ablation procedures. We believe that this method can be practically applied to this and other emerging applications that require histological validation.

WORKS CITED

1. Boaz, T. L., Lewin, J. S., Chung, Y. C., Duerk, J. L., Clampitt, M. E., and Haaga, J. R. MR Monitoring of MR-Guided Radiofrequency Thermal Ablation of Normal Liver in an Animal Model. J.Magn Reson.Imaging 1998;8(1):64-9.

2. Chen, L., Bouley, D. M., Harris, B. T., and Butts, K. MRI Study of Immediate Cell Viability in Focused Ultrasound Lesions in the Rabbit Brain. J.Magn Reson.Imaging 2001;13(1):23-30.

3. Lee, J. D., Lee, J. M., Kim, S. W., Kim, C. S., and Mun, W. S. MR Imaging- Histopathologic Correlation of Radiofrequency Thermal Ablation Lesion in a Rabbit Liver Model: Observation During Acute and Chronic Stages. Korean J Radiol 2001;2:151-8.

4. Merkle, E. M., Haaga, J. R., Duerk, J. L., Jacobs, G. H., Brambs, H. J., and Lewin, J. S. MR Imaging-Guided Radio-Frequency Thermal Ablation in the Pancreas in a Porcine Model With a Modified Clinical C-Arm System. Radiology 1999;213(2):461-7.

5. Merkle, E. M., Shonk, J. R., Zheng, L., Duerk, J. L., and Lewin, J. S. MR Imaging- Guided Radiofrequency Thermal Ablation in the Porcine Brain at 0.2 T. Eur.Radiol. 2001;11:884-92.

6. Morrison, P. R., Jolesz, F. A., Charous, D., Mulkern, R. V., Hushek, S. G., Margolis, R., and Fried, M. P. MRI of Laser-Induced Interstitial Thermal Injury in an in Vivo Animal Liver Model With Histologic Correlation. J.Magn Reson.Imaging 1998;8(1):57-63.

7. Maintz, J. B. and Viergever, M. A. A Survey of Medical Image Registration. Med.Image Anal. 1998;2(1):1-36.

8. Meyer, C. R., Leichtman, G. S., Brunberg, J. A., Wahl, R. L., and Quint, L. E. Simultaneous Usage of Homologous Points, Lines, and Planes for Optimal, 3-D, Linear Registration of Multimodality Imaging Data. IEEE Transactions on Medical Imaging 1995;14(1):1-11.

9. Jacobs, M. A., Windham, J. P., Soltanian-Zadeh, H, Peck, D. J., and Knight, R. A. Registration and Warping of Magnetic Resonance Images to Histological Sections. Med.Phys. 1999;26(8):1568-78.

10. Toga, A. W., Ambach, K., Quinn, B., Hutchin, M., and Burton, J. S. Postmortem Anatomy From Cryosectioned Whole Human Brain. J.Neurosci.Methods 1994;54(2):239-52.

11. Besl, P. J. and McKay, H. D. A Method for Registration of 3-D Shapes. IEEE Transactions on Pattern Analysis and Machine Intelligence 1992;14(2):239-56.

12. Lazebnik, R. S., Lancaster, T. L., Breen, M. S., Lewin, J. S., and Wilson, D. L. Volume Registration Using Needle Paths and Point Landmarks for Evaluation of Interventional MRI Treatments. IEEE Trans.Med.Imaging 2003;22(5):653-60.

13. Bookstein, F. L. Thin-Plate Splines and the Decomposition of Deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence 1989;11(6):567-85.

14. Davis, M. H., Khotanzad, A., Flamig, D. P., and Harris, B. T. A Physics-Based Coordinate Transformation for 3-D Image Matching. IEEE Trans.Med.Imaging 1997;16(3):317-28.

15. Falcao, A. X., Udupa, J. K., Samarasekera, S., Sharma, S., Hirsch, B. E., and Lotufo, D. A. User-Steered Image Segmentation Paradigms: Live Wire and Live Lane. Graphical Models and Image Processing 1998;60(4):233-60.

16. Lancaster, T. L and Wilson, D. L. Correcting Spatial Distortion of Histological Images. Annals of Biomedical Engineering 2003;(In Press).

17. Lazebnik, R. S., Weinberg, B. D., Breen, M. S., Lewin, J. S., and Wilson, D. L. Three-Dimensional Model of Lesion Geometry for Evaluation of MR-Guided Thermal Ablation Therapy. Acad.Radiol 2002;9(10):1128-38.

18. Likar, B. and Pernus, F. Registration of Serial Transverse Sections of Muscle Fibers. Cytometry 1999;37(2):93-106.

19. Mega, M. S, Chen, S. S, Thompson, P. M., Woods, R. P., Karaca, T. J., Tiwari, A., Vinters, H. V., Small, G. W, and Toga, A. W. Mapping Histology to Metabolism: Coregistration of Stained Whole-Brain Sections to Premortem PET in Alzheimer's Disease. Neuroimage 1997;5(2):147-53.

20. Lo, H. W., Tsai, Y. J., Chen, P. H., Chen, H. Y., Ker, C. G., and Juan, C. C. Radiofrequency Ablation for Treatment of Hepatocellular Carcinoma With Cirrhosis. Hepatogastroenterology 2003;50(51):645-50.

21. Schenck, J. F., Jolesz, F. A., Roemer, P. B., Cline, H. E., Lorensen, W. E., Kikinis, R., Silverman, S. G., Hardy, C. J., Barber, W. D., Laskaris, E. T., and . Superconducting Open-Configuration MR Imaging System for Image-Guided Therapy. Radiology 1995;195(3):805-14.

22. Lewin, J. S., Duerk, J. L., Jain, V. R., Petersilge, C. A., Chao, C. P., and Haaga, J. R. Needle Localization in MR-Guided Biopsy and Aspiration: Effects of Field Strength, Sequence Design, and Magnetic Field Orientation. AJR Am.J.Roentgenol. 1996;166(6):1337-45.

23. Liu, H., Hall, W. A., Martin, A. J., and Truwit, C. L. Biopsy Needle Tip Artifact in MR-Guided Neurosurgery. J.Magn Reson.Imaging 2001;13(1):16-22.

24. Cutts, A. Shrinkage of Muscle Fibres During the Fixation of Cadaveric Tissue. J.Anat. 1988;160:75-8.

25. Lazebnik, R. S., Weinberg, B. D., Breen, M. S., Lewin, J. S., and Wilson, D. L. Sub-Acute Changes in Lesion Conspicuity and Geometry Following MR-Guided Radiofrequency Ablation. J Magn Reson.Imaging 2003;18(3):353-9.

100

CHAPTER 5

Laser Thermal Ablation Therapy: Model and Parameter Estimates to Predict Cell Death from MR Thermometry Images

5.1 INTRODUCTION

Solid tumors and other pathologies can be treated using laser thermal ablation under interventional magnetic resonance (MR) image guidance. MR imaging (MRI) has several advantages including the lack of ionizing radiation, excellent soft tissue discrimination, sensitivity to blood flow and temperature, and ability to image at any angle. To monitor an ablation procedure, MRI can continuously acquire temperature images during heating, and structural images during and after heating. We are investigating the ability to monitor treatment using MR thermometry measurements. A model relating temperature history to cell death could be used to predict the therapeutic region in real-time during the heating process, thereby allowing one to treat the pathology and spare adjoining critical tissues.

The use of MR thermometry measurements and a cell death model to predict therapeutic regions has advantages. First, although post-ablation MR lesion images accurately predict the region of cell death (1-4), these can only be obtained several minutes after the ablation, possibly after undesirable damage is done. Because MR temperature images could provide real-time feedback during the ablation, one can cease application at the appropriate moment. Second, post-ablation MR structural images may not distinguish the edema that surrounds the thermal lesion from tumors, hemorrhage, or

101

prior inflammation. Thermometry measurements might be less ambiguous in these cases.

Third, the MR temperature measurements using the proton resonance frequency (PRF) method are accurate and, except for fat, independent of tissue type (5;6). Furthermore, a model of tissue damage is needed in any quantitative pre-procedural planning strategy (7-

9).

There are various potential tissue damage models to use with MR temperature data.

Previous reports have used a critical temperature model that assumes the cell death is not observable below the critical temperature and occurs extremely rapidly and completely above the critical temperature (5;10-15). This model neglects the temperature history.

Others used a mathematical model of the temperature-time relationship for tissue damage. These models usually consisted of either a generalized Arrhenius function

(16;17) or a linear approximation of the Arrhenius function near 43ºC with fixed model parameters (10-13;18;19). However, the empirically-derived values for these Arrhenius- based models can vary for different tissues, temperature ranges, and heating durations

(20).

There are various methods to compare model results to the tissue response. In previous reports, analysis was often done using geometric measurements without alignment of MR temperature images with images of the necrotic region as determined by histology or post-ablation MR data (15-18). For example, Hazle et al. measured diameters of necrotic regions in macroscopic tissue images and calculated diameters of cell death regions predicted by a tissue damage model. In 15 lesions, mean diameters agreed within 2 mm. Obviously, a voxel-by-voxel assessment would be desirable, especially for nonsymmetrical lesions.

102

We developed a new tissue damage model and applied it on a voxel-by-voxel basis to temperature-time histories as obtained from multiple MR thermometry measurements.

Our method includes image registration, image filtering, and iterative parameter estimation techniques. We first describe the model and then apply it to the image data from in vivo rabbit brain.

5.2 MATERIALS AND METHODS

5.2.1 Cell Death Model

We used a mathematical model that predicts the region of cell death. Below, we describe the thermal damage model, fitting procedure, and pertinent mathematics. The complete method and subsequent analysis were implemented using MATLAB software

(Mathworks, Natick, MA).

Our model is based on the local time-varying temperature. As the temperature increases during a thermal ablation, the kinetic energy of the molecules increases. This thermal agitation of molecules at high kinetic energies disrupts chemical bonds which can lead to various destructive cell processes including disintegration of cell membrane bilayers, denaturation of cellular proteins, deactivation of enzymes, and damage of ion channels, subcellular organelles, nucleoplasm, and DNA to cause cell death (21-32). Our assumption is that the severity of these destructive events can increase to a threshold value that leads to cell destruction.

At elevated temperatures, we assume that a normal cell will accumulate “destruction” to a point where it will die. Extending this concept to a slightly more macroscopic view, we assume that there will be destruction (D) to tissue in a region starting in the native

103

condition (N). The build up of destruction will depend both upon temperature and

duration through a temperature dependent rate coefficient, β(T(t)), as shown below.

β(T(t)) N(t) →D(t)

The accumulated destruction of tissue, Ω(τ), at time, τ, is normalized between 0 and 1,

and defined by:

 τ  Ω=(τ)1−exp−∫ β(T(t))dt (5.1)  0 

The temperature response of a heated voxel during ablation will typically show the temperature increase to a maximum, remain at the maximum for some duration, and then decrease to the basal temperature when heating stops. Above a critical temperature, TC,

the destruction of tissue can occur. When the temperature falls below TC, destruction stops. In addition, since the rate of tissue destruction is expected to increase with temperature, β is expected to be a monotonically increasing function with respect to temperature. We used the following mathematical expression for β, which consisted of three parameters (A, N, TC).

0, T(t) < Tc β[T(t)] =  N (5.2) A(T(t)-Tcc) , T(t) ≥ T

The severity of the accumulated destruction increases to a critical threshold value, ΩC, that leads to cell death. The model outputs (M) are the final cell state:

0, Ω(τ ) <Ωc M =  (5.3) 1, Ω(τ ) ≥Ωc

104

where a normal or dead cell state is 0 or 1, respectively. This threshold is based on an all

or nothing tissue response observed in histology minutes post-ablation that shows a sharp

transition between dead and adjacent normal cells (1;33). Thus, the final model includes

four parameters: A, N, TC, ΩC. We will refer to this set of parameters as θ.

5.2.2 Parameter Estimation

We want to fit the model to a known cell state (S), as verified from histology:

0, normal cell S =  (5.4) 1, dead cell

Model outputs, M(θ), are estimated from each voxel in a sequence of temperature

images, T(t), to create an estimated cell death binary image. From the temperature

images, we computed the time integral for Equation 5.1 numerically using trapezoidal

integration. We determine the difference, Dij, between estimated and known cell death images for the ith ablation and jth voxels:

DMij = ij(θ) - Sij, for i=1,...,R, j=1,...,V (5.5)

The sum of voxels incorrectly identified as dead (false-positive, FP), NFP, and normal

(false-negative, FN), NFN, is calculated:

RV NFP = ∑∑ID (5.6) i=1 j=1

1, if D ij = 1 where ID =  0, otherwise

RV NFN = ∑∑KD (5.7) i=1 j=1

105

1, if D ij = −1 where KD =  0, otherwise

We include an adjustable weight, W, to the objective function, f, such that we can place more or less emphasis on NFP or NFN :

f(=W)NFP +(1−W)NFN (5.8)

W will range from 0 to 1. For cancer therapy, we would desire a model that slightly

underestimates the necrotic region (eliminates all FP) to maximize destruction of

malignant cells within the predicted region of necrosis with minimal or acceptable normal

tissue damage. For this case, a large W would be used to emphasize NFP in objective

function. For benign tumor eradication, a lower W would be desirable to achieve a

balance between tumor and normal tissue destruction. For this study, we estimated four

sets of model parameters using W = 0.50, 0.70, 0.90, and 0.99, and investigated the

model performance.

Iterative minimization of Equation 5.8 yields the parameter estimates that best fit the

model to known cell death data. Many optimization algorithms are suitable for a

nonlinear search of parameter space. Typically the choice of algorithm depends on the

tradeoff between the number of objective function evaluations and resistance to local minima. Because our objective function is computationally inexpensive, such that each evaluation requires only a fraction of a second on a Pentium 4 class computer, we used the Nelder-Mead simplex method (34) for its robustness and resistance to discontinuities in parameter space despite a large number of expected iterations.

Initial estimates of the model parameters are determined from computer simulations.

We simulate temperature profiles based on previous studies that showed broad

106

temperature ranges, and heating durations for cell death (23;25;35-37). The model is fitted to the data using an exhaustive search of the parameter space. The resulting set of parameters is the initial set of parameters for the minimization of Equation 5.6.

In order to improve robustness, we add some modifications to the procedure. We perform several passes of optimization. In each pass, the parameter search is paused when an iteration does not reduce f by at least 10-3 of its previous best value. We then perturb the current set of parameters by a random displacement less than 1% of their value and repeat until no further reduction in the objective function is observed following three consecutive searches.

5.2.3 Calibration of PRF Thermal Coefficient

Following a protocol approved by the Administrative Panel on Laboratory Animal Care at Stanford University, we anesthetized six New Zealand white rabbits (3.5-4.0 kg) with a subcutaneous injection of ketamine (35 mg/kg) and xylazine (5 mg/kg). After induction, the rabbits were intubated and anesthesia was maintained with isoflurane (2% to 3%) and oxygen (1%). After surgically removing a piece of each animal’s skull (approximately

25 x 25 mm) and carefully rupturing the dura mater without injury to the brain, we placed the rabbits in the prone position within the bore of a clinical 0.5 T open MR imaging system (GE Medical Systems Signa SP, Milwaukee, WI). A warm plastic blanket was placed between the rabbit and the scanner table to aid in thermoregulation. We placed an extremity coil above the brain.

We inserted two fiberoptic temperature sensors (Luxtron Corporation, Santa Clara,

CA) 8 mm apart and 8 mm deep into the brain. A customized 30 feet long, 0.6 mm

107

diameter laser fiber with a 5 mm long, 0.2 mm diameter conical laser diffusion tip

(Surgical Laser Technologies, Montgomeryville, PA) was inserted 8 mm into the brain and between the two temperature sensors. All three fibers were positioned in one hemisphere in a sagittal MRI plane, approximately 3 mm lateral to the longitudinal fissure. To confirm the alignment of the two temperature sensors with the MRI scan plane, we applied a T1-weighted fast spin-echo (FSE) sequence with TR/TE parameters of 500/17 msec that gives 256 x 128 voxels over a 16 x 16 cm FOV and 3.0 mm slice thickness to yield 0.625 x 1.25 x 3.0 mm voxels. Using a 2 W laser power source operating at a wavelength of 1064 nm and located outside the scan room, we applied laser energy to increase local tissue temperature via light energy absorption for heating durations between 1 and 8 minutes. The maximum temperatures measured by the temperature sensors ranged from 45°C to 85°C (∆T = 10 - 47°C).

Before, during, and after heating, we continuously acquired MR images of the brain.

With three TE settings in separate experiments, a gradient-echo (GE) sequence was applied with TR/TE parameters of 77.2/38.9, 21.5, 13.3 msec, flip angle of 30°, image time of 10 sec, that gives 256 x 128 voxels over a 16 x 16 cm FOV and 3.0 mm slice thickness to yield 0.625 x 1.250 x 3.0 mm voxels. A reference phantom, separated from the animal by 2 cm, was used to correct any scalar phase drift during imaging.

Immediately after repeating the procedure on the other hemisphere, the rabbits were euthanized via intravenous administration of a euthanasia solution (1 ml/4.5 kg).

We used a 2 x 2 voxel region of interest (ROI) for the MR temperature measurement at the tip of the thermal sensor. Paired data (MR, fiberoptic sensor) were obtained from

1140 time-points. Lines were fitted to the phase shift (∆φ) vs. temperature elevation (∆T)

108

for each TE, and to ∆φ/∆T vs. TE, respectively. The thermal coefficient was found from the fitted line. The data analysis indicated an interaction between the laser beam and the fiberoptic temperature sensors. As a result, our calibrations only used the data points acquired when the laser was turned off.

5.2.4 Laser Ablation and MR Imaging

To correlate the model predicted regions of cell death with the tissue response, we created a thermal lesion in seven rabbit brains. Each animal was anesthetized and prepared as described above. We placed the rabbit in the bore of the open MRI system in the prone position. The laser fiber was inserted approximately 8 mm into each rabbit brain and 3 mm lateral to the longitudinal fissure. Two rice noodles (approximately 1 mm in diameter) were inserted 8 mm into each brain, with one on each side and parallel to the laser fiber. The laser fiber and the two rice noodles were located within a sagittal MR scan plane. Using a fiberoptic temperature sensor that was inserted 8 mm into each brain and approximately 4 mm lateral to the laser fiber, we measured the baseline temperature.

To correct any phase drift during imaging, we used a reference phantom which was separated from the animal by 2 cm. To verify the alignment of the two fiducials with the

MRI scan plane, a T1-weighted FSE sequence was applied with the same parameters as calibration image acquisition described above.

Lesion formation was achieved by increasing the local tissue temperature with the absorption of light energy delivered from the laser diffusion tip. Using the 2 W laser power source located outside the scan room, we applied laser energy for various heating durations as shown in Table 1. To confirm the formation of a thermal lesion in each

109

experiment, heating was applied until a temperature of at least 55°C was measured by the temperature sensor.

Before, during, and after heating, we continuously applied a gradient-echo (GE) sequence with TR/TE parameters of 77.2/38.9 msec, flip angle of 30°, image time of

10 sec, that gives 256 x 128 voxels over a 16 x 16 cm FOV and 3.0 mm slice thickness to yield 0.62 x 1.25 x 3.0 mm voxels. Based on our previous studies, in which the thermal lesion in MR images correlated well with the cell death zone at four hours post-ablation

(3), the rabbits were kept in the magnet under general anesthesia for four hours. At four hours post-ablation, we applied a T2-weighted spin-echo (SE) MR images in the same orientation as the GE MR images with TR/TE parameters of 4000/115 msec, number of excitations = 4, that gives 512 x 256 voxels over a 16 x 16 cm FOV and 2.0 mm slice thickness to yield voxels = 0.31 x 0.62 x 2.0 mm. The rabbits were then euthanized using the same method as described above.

5.2.5 Processing of MR Temperature Maps

To provide immediate assessment of the heating, MR temperature maps were processed in near real-time. We subtracted pre-ablation baseline phase maps and from remaining phase maps by multiplying each image by the complex conjugate of the baseline image on a voxel-by-voxel basis. To correct any scalar phase drift during heating, the phase measured in the separate phantom was similarly subtracted from the phase maps. This processing was repeated retrospectively at the time of image analysis.

110

5.2.6 Temporal Filtering of MR Temperature Image Sequence

Since GE MR phase images can be noisy on a low-field system, we applied temporal filtering to the temperature maps. We removed transient noise spikes with a temporal grayscale morphological filter. Each voxel was temporally filtered across the entire sequence of temperature maps. Using a flat disc-shaped structuring element with a one pixel radius, we performed a morphological opening and then closing operation to remove positive and negative temperature noise spikes, respectively. Movies of the filtered temperature map sequences were created to visually confirm the elimination of the transient noise spikes. This filtering process was implemented in a conservative manner to produce an enhanced image sequence with reduced noise, while leaving the overall temperatures relatively undisturbed.

5.2.7 Registration of Post-ablation to Temperature MR Images

Before aligning the post-ablation T2-weighted MR lesion image with the sequence of

GE MR images used for temperature mapping, we performed some preprocessing steps.

First, we created movies of the sequence of GE MR images to visually confirm that there was no noticeable motion during the ablation. To improve the signal to noise ratio of the

GE MR image used for registration, we averaged 16 pre-ablation images. Finally, we removed the non-rigid neck region in the images.

Using Analyze, we performed a two dimensional (2D) rigid body registration to align fiducials and the outer boundary of the skull in post-ablation MR image to the reference

GE MR image. To visually evaluate alignment of images, the reference and transformed images were displayed with a linked digitally imposed crosshair that allowed us to

111

compare locations within the images. Also, a fused image was shown that overlayed the grayscale intensities from one image to a color map of the other. With real-time display of the transformed and fused images, we interactively adjusted the rigid body parameters with a cursor and evaluated the registration. We outlined the fiducials and skull boundary in the reference image, and copied them to the transformed image. The optimal parameters were used to transform the post-ablation MR image with bicubic interpolation.

5.2.8 Segmentation of Cell Death Region

To compare the model predicted regions of cell death with the tissue response, we created a binary cell death image. After registering the MR images, we manually segmented the outer boundary of the hyperintense rim in the post-ablation T2-weighted MR lesion images. The correlation of the outer boundary of the hyperintense rim in hyperacute and subacute T2-weighted MR lesion images with the region of necrosis assayed from registered histology, was previously validated (1-4;33).

One of the authors (MB) performed the segmentation using a region of interest (ROI) software program written in MATLAB. The observer had experience looking at MR thermal lesion images without knowledge of modeled regions of cell death. The investigator previously established criteria for lesion boundaries under the supervision of a radiologist specializing in interventional MRI thermal ablation. The observer followed a strict protocol. All segmentation was performed on the same workstation in a darkened room. We perceptually linearized the display using Optical (ColorVision Optical,

Rochester, NY) and validated with a step wedge image. By adjusting the vertical and

112

horizontal controls of the display, we obtained square pixels with 0.2 mm edge length.

Window and level settings for each image were fixed prior to segmentation such that the qualitative contrast between the lesion’s hyperintense rim and adjacent normal tissue was maximized.

After segmentation, each voxel value along and inside the segmented boundary was set to one, else it was zero. The set of segmented binary cell death images from all experiments was used for subsequent model fitting and error estimation.

5.2.9 Evaluation of Model Performance

To quantitatively evaluate the tissue damage model fit, we directly compared the model output to a manually segmented cell death binary image. For each voxel, our model calculates the accumulated destruction, Ω, and the final cell state, M, as shown in

Equation 5.1 and 5.3, respectively. Hence, we assessed the quality of model fit for both outputs.

To evaluate the model fit accuracy for Ω, we used receiver operating characteristic

(ROC) curve analysis [Metz, Swetz]. For each ablation, we created an image of Ω, and calculated the ROC curve and area under the ROC curve (Az) to quantitatively assess the model’s ability to discriminate between regions of dead and normal cells, as determined from the S image. In the Ω image, pixel values near one and zero should correspond to regions of dead and normal cells, respectively. A threshold was applied to each Ω image at 100 levels with equal increments to create segmentations where pixels above and below the threshold are marked as dead and normal, respectively. At each threshold, we obtained the number of voxels correctly identified as positive (true positive, TP) and

113

negative (true negative, TN), and incorrectly identified as positive (false positive, FP) and negative (false negative, FN), and computed a true-positive fraction (TPF) and a false-positive fraction (FPF):

number of pixels correctly determined as dead TP TPF == (5.9) total number of dead pixels TP + FN

number of pixels incorrectly determined as dead FP FPF == (5.10) total number of normal pixels TN + FP

To create an ROC curve, we plotted the TPF as a function of FPF for the various thresholds. Since we computed the FPF and TPF from a large number of samples with small increments in threshold, we obtained relatively smooth curves, and computed the area, Az, numerically using trapezoidal integration. For each thermal lesion, the ROC curve was plotted, and Az was estimated.

To assess the model fit error for M, we directly compared the model and segmented binary cell death images. We determined the number of TP, FP, TN, and FN, and computed the sensitivity and specificity. We created a difference image with each voxel assigned a color-coded value of 0, 1, 2, and 3 for TP, FP, TN, and FN, respectively. We also calculated the distance from each mislabeled voxel to the segmented border of cell death. Each voxel was assigned a 2D coordinate, in millimeters, based on the pixel size.

A continuous spline was interpolated along the manually segmented cell death border.

For each mislabeled voxel, an automatic algorithm found the closest point along the spline of the cell death border, and the Euclidean distance between these points was calculated.

We compared our cell death model to the critical temperature model. With this model, thermal damage occurs if tissue exceeds a threshold temperature, with no damage

114

below the threshold limit. To compare the model errors, we calculated the difference in the number of FP and FN across all lesions for each W.

5.3 RESULTS

In this study, we found the phase change, ∆φ/∆T (in degrees/°C), as a function of TE by

∆∆φ T =−0.0098γΒο TE +0.48 (5.11) where γ is the gyromagnetic ratio, Bo is the magnetic field strength, and 0.48 a phase offset.

Figure 5.1 shows typical MR temperature maps acquired during a laser ablation of rabbit brain. The maps show the increasing intensity and spatial distribution of temperature over time. The heated region has an elliptical temperature profile since the laser fiber is parallel to the imaging plane. The low level of noise in the images is evident with only small temperature changes seen in the unheated peripheral region. The heating duration and maximum temperature for each ablation are shown in Table 5.1.

115

Figure 5.1. MR temperature maps acquired during a typical laser ablation of rabbit brain. The maps were acquired at 0.5 (A), 1.0 (B), 2.0 (C), and 5.0 (D) minutes during heating. The MR imaging plane was oriented parallel to the laser fiber. Image acquisition time was 10 sec. The images were temporally filtered with a grayscale morphological filter to remove noise spikes.

The post-ablation T2-weighted MR images were well registered with the GE MR

images used for temperature mapping. In Figure 5.2, fiducials were marked with

graphical overlays in the GE MR image and copied to the registered post-ablation MR

image, where they were well aligned with corresponding fiducials. Similar results were

obtained for the other rabbit ablation experiments.

In Table 5.2, the estimated parameters for our cell death model are shown for each

W. Since the parameters were optimized simultaneously using all lesions, there is one set

of estimated parameters for each W. For each MR temperature map, the time to solve the

model equations on a Pentium 4 class computer was 16 ms for a 25 x 25 voxel region of

interest.

116

Figure 5.2. Registration quality of GE MR images used for temperature mapping and MR lesion images. Images are: T2-weighted MR lesion image (top) acquired four hours p ost-ablation and GE MR image averaged from 16 pre-ablation images. Averaging the GE MR images significantly improved the signal to noise ratio. In the post-ablation MR im age, the thermal lesion (vertical arrow) is the elliptical region that has a isointense central core surrounded by a hyperintense rim. The two rice noodle fiducials are thin dark vertical lines in the MR images, which were inserted on each side of a laser fiber before heating for image registration. The fiducials were marked with horizontal arrows in the GE MR image and copied to the registered post-ablation MR image. Excellent registration accuracy of MR images is clearly evident with good correspondence of fiducials.

117

We examined the model fit error for the accumulated tissue damage and final cell

state. In Figure 5.3, an overlay of the segmented cell death boundary in the registered

post-ablation MR image was copied to the map of the modeled tissue damage for

lesion 3. The modeled tissue damage is typically much greater inside the cell death

boundary than outside, indicating a good model fit. To assess this quantitatively, we

calculated ROC curves for each lesion. In Figure 5.4, we plotted, as a function of lesion,

the area under the ROC curves. These areas compared favorably to 1.0, the best possible

value. For all lesions combined, the area under the ROC curve was 0.96, 0.96, 0.88, and

0.77 for W of 0.50, 0.70, 0.90, and 0.99, respectively.

Figure 5.3. Comparison of segmented cell death boundary with modeled tissue damage map. The same T2-weighted post-ablation MR (left) and model accumulated tissue damage (righ t) images are shown in each row with overlays displayed only at the bottom. Based on previous studies, the outer boundary of the hyperintense rim in the post-ablation T2-weighted MR image corresponds to boundary of necrosis as seen in registered histology images on the order of one MR voxel. The cell death boundary was manually segmented in the registered T2-weighted MR lesion image and copied to the color-coded tissue damage map.

118

Figure 5.4. Plotted are the areas under the ROC curves for each thermal lesion. The lesions are ordered by heating duration. The ROC curves were calculated after parameters were optimized simultaneously for all lesions. Results are shown for a W of 0.5. All areas were greater than 0.94, with a maximum area of 0.99 obtained from lesion 2 and 4. Across all lesions, the area was 0.96, a value close to 1.0.

To evaluate the model fit error for the final cell state, we compared the model and

segmented binary cell death images. In Figure 5.5, we showed a difference map for

lesion 3 for a W of 0.5. The same lesion was used in Figure 5.3. There were no FP. FN

were typically within one pixel of the segmented cell death boundary, indicating a slight

underestimation of the necrotic region.

To further evaluate this quantitatively, we calculated the number of model errors. In

Figure 5.6, we plotted as a function of lesion, the number of FP and FN for a W of 0.50.

The number of FP was small as compared to the lesion size. In addition, the number of

FN was typically more than the number of FP, indicating the model’s tendency to

underestimate the necrotic region. In Figure 5.7, we plotted as a function of W, the

number of FP and FN across all lesions. For all lesions, a total of 4375 voxels were

analyzed that included 766 lesion voxels and 3609 normal voxels. The minimum number

119

of total errors was obtained at a W of 0.50. As W increased, the number of FP and FN

increased and decreased, respectively. At a W of 0.99, the model always underestimates

the cell death region since there was no FP.

Figure 5.5. Difference map of the model fit error for a typical ablation. The error map was obtained after parameters were estimated using data in a 25 x 25 voxel region of interest from all lesions simultaneously. Each voxel, with in-plane dimensions of 0.625 x 0.625 mm, was color-coded as either TP, TN, FP, and FN. FP and FN indicate an overestimation and underestim ation of the necrotic region. For this ablation, only voxels along the boundary of the cell death region were mislabeled as FN, with no FP.

In Figure 5.8, we plotted the specificity and sensitivity across all lesion as a function

of W. With larger W, the specificity and sensitivity increased and decreased, respectively.

The specificity was consistently above 0.98, a value close to 1.0, indicating only a slight overestimation of the cell death region.

120

Figure 5.6. Plotted are the lesion size, number of FP, and number of FN for each thermal lesion. The lesions are ordered by heating duration. At an objective function weight of 0.50, the model parameters were estimated simultaneously for all lesions. The FP and FN correspond to overestimation and underestimation of the cell death region, respectively. For lesion 1 and 3, there were no FP. The relatively small size of lesion 4 is probably due to insufficient penetration of laser energy into the tissue.

Figure 5.7. Plot of the number of FP and FN across all lesions as a function of W. The size the ac tual cell death region (766 voxels) is indicated with a line. At each W, we estimated parameters simultaneously for all lesions. There are no FP at W equal to 0.99. The FP and FN correspond to overestimation and underestimation of the cell death region, respectively.

121

Figure 5.8. Plot of the specificity and sensitivity across all lesions as a function of W. At each W, we estimated parameters simultaneously for all lesions. At W equal to 0.99, the specificity is perfect, 1.0, since there are no FP. Specificities and sensitivities less than 1.0 correspond to overestimations and underestimations of the cell death region, respectively.

In Figure 5.9, we evaluated the model fit by plotting the median distance between

model errors and the segmented cell death boundary as a function of W. We chose the

median statistic rather than the mean because the distribution of the distances is not

Gaussian or symmetric about the mean. The median distances for the FP and FN were

always less or equal to 0.85 mm and 0.34 mm, respectively, which agrees favorably with

the in-plane voxel dimensions of 0.62 x 1.25 mm.

We also compared our model with the critical temperature cell death model. In

Figure 5.10, we plotted the difference in the number of FP and FN between each model

as a function of W. At a W of 0.99, both models produced no FP. For all other cases, our

122

model predicted fewer errors. A maximum difference of 13 FP and 57 FN was observed

at a W of 0.50 and 0.90, respectively.

Figure 5.9. Plot of median distance between segmented cell death boundary and FP and FN voxels as a function of W. Each bar represents the median, 25 and 75 percentiles across all lesions. FP and FN voxels are outside and inside the segmented cell death boundary, respectively. For a W of 0.99, there are no FP.

Figure 5.10. Plot of the difference in number of FP and FN between our model and the critical temperature model as a function of W. For the units on the right axis, the number of cells in a voxel was based on a cell size of 20um. For both models, the minimum number of total errors was observed at a W of 0.50.

123

5.4 DISCUSSION

Our results suggest that it is possible to use in vivo MR temperature images and a thermal

damage model to predict regions of cell death. Because of its simplicity, our model could

be practically applied during thermal ablations. In real-time, model predicted regions of cell death could be overlaid on iMRI images or even on registered single photon emission computed tomography (SPECT) and high resolution MR tumor images acquired before treatment. Predictions could then be visually compared to a segmented pathology to give

real-time feedback to the physician. Based on our different sets of parameter estimates,

the model could be adjusted to place more or less emphasis on tumor versus normal

tissue destruction. Both the preprocessing time and the time to solve the equations are

negligible as compared to the typical time of 10 sec. between MR thermometry images.

The same model should be applicable to other potential temperature measuring

techniques such as ultrasound.

By selecting various parameter estimates, our model could be tailored to the clinical

treatment. For cancer therapy in organs with regenerative properties, such as the liver, we would desire a model that slightly underestimates the necrotic region to maximize destruction of tumor cells with acceptable damage to surrounding normal tissue. For

benign tumors, a different model could be used to achieve a balance between tumor and

normal tissue destruction.

We can compare our model and data analysis technique to previously reported ones.

Arrhenius-based models using parameters from other experiments not surprisingly did not always work (10). A time-temperature product model did not work (11). Maximum temperature typically does much better (10;11). We use a model that considers the

124

temperature time history. In principal this model will be able to predict a wider range of

temperature time histories. The latter point is being investigated. Our analysis method

uses all interesting voxels unlike previous reports which only analyzed voxels along the

segmented cell death boundary (10-12). Hence, in theory, we can more precisely assess

the model fit error.

There are several important aspects to our method. Features such as accurate image

registration, filtering, and parameter optimization are important steps to accurately fit the

model to the segmented region of cell death. Previous studies comparing modeled tissue

damage from MR temperature maps with the tissue response used models and parameter

values derived from experiments at lower temperatures and significantly longer heating

durations than typical clinical ablations (20). Our model with optimized parameters and

analysis methodology on a voxel-by-voxel basis addresses these limitations and should

enable a more accurate prediction of the necrotic region.

Even though our model validation was based on thermal ablations of normal brain

tissue, we expect similar results in tumor and other tissue types. As the temperature

increases during a thermal ablation, tissue damage occurs due to excessive physical

forces between molecules which can lead to various destructive cell processes such as rupturing of cell membranes, denaturation of proteins, and damage of ion channels (21-

31). Hence, the tissue’s vulnerability to thermal damage depends on molecular structure and strength of chemical bonds but not on cell size or shape. Therefore, the susceptibility

to heat damage is quite similar across different tissue types.

The importance of this technique can be seen in its application to determine the

tissue damage and destruction in real time during actual treatment, in order that treatment

125

could be adjusted for temperature variations during the actual treatment. For example, in a case where the measured temperature exceeded the proposed treatment, the actual treatment could be shortened to correct for the extra damage occurring during the period of excessive temperature. Alternatively, in some instances, it may not be possible to achieve a pre-chosen temperature because of such limitations as insufficient power, tumor location, high tumor blood flow, or a combination of these factors. The ability to monitor therapy in real-time could extend the method to the safe destruction of tumor adjacent to vital structures that might be damaged by heating such as the gall bladder, bowel, and especially the brain, where collateral damage must be minimized.

We conclude that our tissue damage model coupled with a sequence of MR temperature maps can be used to accurately predict the tissue response. Results show that for in vivo rabbit brain, the estimated region of necrosis closely corresponds to the segmented region of cell death. This is good evidence that MR temperature maps can be used with our thermal damage model to predict the therapeutic region in real-time.

126

WORKS CITED

1. Breen, M. S., Lazebnik, R. S., Fitzmaurice, M., Nour, S. G., Lewin, J. S., and Wilson, D. L. Radio-Frequency Thermal Ablation: Correlation of Hyperacute MR Lesion Images With Tissue Response. J Magn Reson.Imaging 2004;Submitted.

2. Lazebnik, R. S., Breen, M. S., Fitzmaurice, M., Nour, S. G., Lewin, J. S., and Wilson, D. L. Radio-Frequency-Induced Thermal Lesions: Subacute Magnetic Resonance Appearance and Histological Correlation. J.Magn Reson.Imaging 2003;18(4):487-95.

3. Chen, L., Bouley, D. M., Harris, B. T., and Butts, K. MRI Study of Immediate Cell Viability in Focused Ultrasound Lesions in the Rabbit Brain. J.Magn Reson.Imaging 2001;13(1):23-30.

4. Morrison, P. R., Jolesz, F. A., Charous, D., Mulkern, R. V., Hushek, S. G., Margolis, R., and Fried, M. P. MRI of Laser-Induced Interstitial Thermal Injury in an in Vivo Animal Liver Model With Histologic Correlation. J.Magn Reson.Imaging 1998;8(1):57-63.

5. Kahn, T., Harth, T., Kiwit, J. C., Schwarzmaier, H. J., Wald, C., and Modder, U. In Vivo MRI Thermometry Using a Phase-Sensitive Sequence: Preliminary Experience During MRI-Guided Laser-Induced Interstitial Thermotherapy of Brain Tumors. J Magn Reson.Imaging 1998;8(1):160-4.

6. Stollberger, R., Ascher, P. W., Huber, D., Renhart, W., Radner, H., and Ebner, F. Temperature Monitoring of Interstitial Thermal Tissue Coagulation Using MR Phase Images. J Magn Reson.Imaging 1998;8(1):188-96.

7. Johnson, P. C. and Saidel, G. M. Thermal Model for Fast Simulation During Magnetic Resonance Imaging Guidance of Radio Frequency Tumor Ablation. Ann.Biomed.Eng 2002;30(9):1152-61.

8. Puccini, S., Bar, N. K., Bublat, M., Kahn, T., and Busse, H. Simulations of Thermal Tissue Coagulation and Their Value for the Planning and Monitoring of Laser- Induced Interstitial Thermotherapy (LITT). Magn Reson.Med. 2003;49(2):351-62.

9. Hayes, L. J., Diller, K. R., Pearce, J. A., Schick, M. R., and Colvin, D. P. Prediction of Transient Temperature Fields and Cumulative Tissue Destruction for Radio Frequency Heating of a Tumor. Med.Phys. 1985;12(6):684-92.

10. Chen, L., Wansapura, J. P., Heit, G., and Butts, K. Study of Laser Ablation in the in Vivo Rabbit Brain With MR Thermometry. J Magn Reson.Imaging 2002;16(2):147- 52.

11. Peters, R. D., Chan, E., Trachtenberg, J., Jothy, S., Kapusta, L., Kucharczyk, W., and Henkelman, R. M. Magnetic Resonance Thermometry for Predicting Thermal

127

Damage: an Application of Interstitial Laser Coagulation in an in Vivo Canine Prostate Model. Magn Reson.Med. 2000;44(6):873-83.

12. Graham, S. J., Chen, L., Leitch, M., Peters, R. D., Bronskill, M. J., Foster, F. S., Henkelman, R. M., and Plewes, D. B. Quantifying Tissue Damage Due to Focused Ultrasound Heating Observed by MRI. Magn Reson.Med. 1999;41(2):321-8.

13. McDannold, N. J., King, R. L., Jolesz, F. A., and Hynynen, K. H. Usefulness of MR Imaging-Derived Thermometry and Dosimetry in Determining the Threshold for Tissue Damage Induced by Thermal Surgery in Rabbits. Radiology 2000;216(2):517- 23.

14. Vykhodtseva, N., Sorrentino, V., Jolesz, F. A., Bronson, R. T., and Hynynen, K. MRI Detection of the Thermal Effects of Focused Ultrasound on the Brain. Ultrasound Med.Biol. 2000;26(5):871-80.

15. Schulze, P. C., Kahn, T., Harth, T., Schwurzmaier, H. J., Schober, R., and Schulze, C. P. Correlation of Neuropathologic Findings and Phase-Based MRI Temperature Maps in Experimental Laser-Induced Interstitial Thermotherapy. J Magn Reson.Imaging 1998;8(1):115-20.

16. Schwarzmaier, H. J., Yaroslavsky, I. V., Yaroslavsky, A. N., Fiedler, V., Ulrich, F., and Kahn, T. Treatment Planning for MRI-Guided Laser-Induced Interstitial Thermotherapy of Brain Tumors--the Role of Blood Perfusion. J Magn Reson.Imaging 1998;8(1):121-7.

17. Sherar, M. D., Moriarty, J. A., Kolios, M. C., Chen, J. C., Peters, R. D., Ang, L. C., Hinks, R. S., Henkelman, R. M., Bronskill, M. J., and Kucharcyk, W. Comparison of Thermal Damage Calculated Using Magnetic Resonance Thermometry, With Magnetic Resonance Imaging Post-Treatment and Histology, After Interstitial Microwave Thermal Therapy of Rabbit Brain. Phys.Med.Biol. 2000;45(12):3563-76.

18. Hazle, J. D., Stafford, R. J., and Price, R. E. Magnetic Resonance Imaging-Guided Focused Ultrasound Thermal Therapy in Experimental Animal Models: Correlation of Ablation Volumes With Pathology in Rabbit Muscle and VX2 Tumors. J Magn Reson.Imaging 2002;15(2):185-94.

19. McDannold, Hynynen, K., Wolf, D., Wolf, G., and Jolesz, F. MRI Evaluation of Thermal Ablation of Tumors With Focused Ultrasound. J Magn Reson.Imaging 1998;8(1):91-100.

20. Sapareto, S. A. and Dewey, W. C. Thermal Dose Determination in Cancer Therapy. Int.J Radiat.Oncol.Biol.Phys. 1984;10(6):787-800.

21. Gershfeld, N. L. and Murayama, M. Thermal Instability of Red Blood Cell Membrane Bilayers: Temperature Dependence of Hemolysis. J Membr.Biol. 1988;101(1):67-72.

128

22. Lee, R. C. and Astumian, R. D. The Physicochemical Basis for Thermal and Non- Thermal 'Burn' Injuries. Burns 1996;22(7):509-19.

23. Thomsen, S. Pathologic Analysis of Photothermal and Photomechanical Effects of Laser-Tissue Interactions. Photochem.Photobiol. 1991;53(6):825-35.

24. Leyko, W. and Bartosz, G. Membrane Effects of Ionizing Radiation and Hyperthermia. Int.J Radiat.Biol.Relat Stud.Phys.Chem.Med. 1986;49(5):743-70.

25. Thomsen, S., Pearce, J. A., and Cheong, W. F. Changes in Birefringence As Markers of Thermal Damage in Tissues. IEEE Trans.Biomed.Eng 1989;36(12):1174-9.

26. Bruneval, P., Mesnildrey, P., and Camilleri, J. P. Nd-YAG Laser-Induced Injury in Dog Myocardium: Optical and Ultrastructural Study of Early Lesions. Eur.Heart J 1987;8(7):785-92.

27. Dewey, W. C. Failla Memorial Lecture. The Search for Critical Cellular Targets Damaged by Heat. Radiat.Res. 1989;120(2):191-204.

28. Coss, R. A., Dewey, W. C., and Bamburg, J. R. Effects of Hyperthermia on Dividing Chinese Hamster Ovary Cells and on Microtubules in Vitro. Cancer Res. 1982;42(3):1059-71.

29. Bischof, J. C., Padanilam, J., Holmes, W. H., Ezzell, R. M., Lee, R. C., Tompkins, R. G., Yarmush, M. L., and Toner, M. Dynamics of Cell Membrane Permeability Changes at Supraphysiological Temperatures. Biophys.J 1995;68(6):2608-14.

30. Borrelli, M. J., Wong, R. S., and Dewey, W. C. A Direct Correlation Between Hyperthermia-Induced Membrane Blebbing and Survival in Synchronous G1 CHO Cells. J Cell Physiol 1986;126(2):181-90.

31. Lepock, J. R. Involvement of Membranes in Cellular Responses to Hyperthermia. Radiat.Res. 1982;92(3):433-8.

32. Roti Roti, J. L. and Winward, R. T. Factors Affecting the Heat-Induced Increase in Protein Content of Chromatin. Radiat.Res. 1980;81(1):138-44.

33. Breen, M. S., Wilson, D. L., Lazebnik, R. S., and Lewin, J. S. Three Dimensional Comparison of Interventional MR Radiofrequency Ablation Images With Tissue Response. Medical Image Computing and Computer-Assisted Intervention, Lecture Notes in Computer Science 2003;2879:24-31.

34. Nelder, J. A. and Mead, R. A Simplex Method for Function Minimization. Computer Journal 1965;7:308-13.

35. Bromer, R. H., Mitchell, J. B., and Soares, N. Response of Human Hematopoietic Precursor Cells (CFUc) to Hyperthermia and Radiation. Cancer Res. 1982;42(4):1261-5.

129

36. Goldberg, S. N., Gazelle, G. S., Halpern, E. F., Rittman, W. J., Mueller, P. R., and Rosenthal, D. I. Radiofrequency Tissue Ablation: Importance of Local Temperature Along the Electrode Tip Exposure in Determining Lesion Shape and Size. Acad.Radiol. 1996;3(3):212-8.

37. Wood, B. J., Ramkaransingh, J. R., Fojo, T., Walther, M. M., and Libutti, S. K. Percutaneous Tumor Ablation With Radiofrequency. Cancer 1-15-2002;94(2):443-51.

130

Chapter 6

Conclusions and Future Work

6.1 Conclusions

In this work, we have developed experimental and computer methods to accurately align

MR images with macroscopic tissue and histology images. Using these registered images, we determined that MR thermal lesion images can be used to predict cell death and damage. We also developed and evaluated a mathematical model that successfully predicts cell death from the temperature history. A summary of our main results and conclusions follows.

In Chapter 2, we developed novel experimental and computer methods to accurately map the tissue response to MR thermal lesion images. Critical experimental developments included the use of intermediate photographs of thick tissue slices obtained using a specially designed apparatus, and methods to minimize tissue deformation and destruction during dissection and slicing processes. Features such as the two dimensional

(2D) warping registration of histology images and three dimensional (3D) needle registration of MR volumes were important steps to accurately align histology to in vivo

MR images. For these experiments, we achieved an overall 3D registration error which compared favorably to the MR voxel dimensions. We believe that we were the first to perform a 3D alignment of in vivo MR lesion images with histology to enable more accurate correlation of MR signal changes with ablation pathology.

In Chapter 3, we evaluated cell death and damage using hyperacute MR thermal lesion images acquired approximately 45 minutes post-ablation. In the hyperacute T2 and

131

CE T1-weighted MR lesion images, our findings strongly suggest that the outer border of

the hyperintense rim corresponds to the region of eventual cell necrosis within a distance less than our ability to measure. Features of our method such as 3D registration of in vivo

MR images to histology images, accurate segmentation of tissue damage boundaries on tiled images of large-format histology slides, and reliable assays to determine tissue damage such as polarized light assessment of muscle birefringence, were important steps to accurately correlate the tissue response to MR thermal lesions images. These findings are good evidence that during RF ablation treatments, MR lesion images can be used to accurately localize the zone of irreversible tissue damage at the lesion margin.

In Chapter 4, to unequivocally reveal the complete extent of tissue necrosis in histology, we performed studies with animals sacrificed four days post-ablation. Using an ellipsoidal model to describe the lesion surfaces, a correlation of histology and

T2-weighted MR images four days post-ablation showed the boundary of cell death

corresponds to the outer border of the hyperintense region with sub-voxel accuracy. We

previously determined that the outer border of the hyperintense region in MR images

four days and 45 minutes post-ablation were not statistically different (1-3). Therefore,

we inferred that the outer border of the hyperintense region in the MR images acquired

45 minutes post-ablation is a reliable marker of cell death with no evidence of cell

recovery.

In Chapter 5, we developed a model and parameter estimates to predict cell death

from the temperature history. We demonstrated that our tissue damage model coupled

with a sequence of MR temperature maps can be used to accurately predict the tissue

response. For laser thermal ablations created in rabbit brain, the estimated region of

132

necrosis closely corresponds to the region of cell death, as determined from histology.

This is good evidence that in vivo MR temperature maps can be used with our thermal damage model to predict the therapeutic region in real-time. We believe that our model, with optimized parameters and analysis on a voxel-by-voxel basis, addresses limitations of previous studies, and should enable a more accurate prediction of the necrotic region.

Overall, our work demonstrated the ability to quantitatively predict tissue damage and death using MR image acquisition and analysis methods, and a mathematical model that accounts for the tissue response to temperature history. With this research, we established an analysis paradigm suitable for many future studies of ablation techniques, localized drug release, and other interventional MRI-guided therapies.

6.2 Future Work

To build on the work presented in this dissertation, we propose future work in five main areas. First is a more detailed examination of the tissue response. Second is an investigation with other tissue types, including tumor, to examine any associated effects on the relationship between MR images and the tissue response. Third is the investigation of alternative ablation techniques, such as laser and focused ultrasound, and image- guided drug delivery therapies. Fourth is the development of improved registration techniques to compare 3D medical images to the tissue response. Finally, we could investigate the use of molecular imaging to monitor therapy.

133

6.2.1 Tissue Response

An important aspect of this research is the careful histological characterization. In our

investigation, we examined the tissue response based on changes in cell morphology, stain color, and birefringence. The determination of cell death is challenging from morphological changes in histology that is fixed as little as one hour following ablation.

Even though we determined that the region of severely damaged tissue minutes after the ablation becomes necrotic four days post-ablation (1-3), there is a slight possibility for

metastasis. Since various enzymes are released in the minutes following a cell injury,

enzyme histochemistry could be used to provide additional confidence and potentially

improved sensitivity for the detection of any surviving cells. Additionally, one could use

triphenyl tetrazolium chloride (TTC), a vital enzyme stain reported to be a reliable

indicator of cell death (4) to assess the cell damage in macroscopic tissue. Enzyme

activity changes could be mapped to observed morphological changes to further our understanding of the tissue response.

In the histologically stained muscle, damage seems all or nothing at the margin, and this phenomenon deserves special consideration. Additional studies could be performed to investigate potential mechanisms. With careful mapping of the tissue response to MR temperature images, one could examine the spatial temperature gradients at the margin. A sharp temperature change at the margin could possibly explain the distinct cell death boundary. Alternatively, biology could lead to an all or nothing response. New information from enzyme activity mapping could provide a spatial distribution of the severity of cell damage at the margin. As a possible mechanism, one

134

could investigate the bystander effect in thermal stress, in which cells at a distance could

be affected by damage to other cells. This effect has been previously shown for radiation

stress to cells but may also apply to thermal stress (5-7). Since it is impossible to

thermally dose cells selectively, an experiment would consist of heating cells, removing

the fluid, and determining if the fluid affects other cells. An improved understanding of

the tissue response might lead to changes in the clinical protocol to improve ablation

procedures.

6.2.2 Tissue Types and Tumor Model

In our ablation experiments, we used a rabbit skeletal muscle model to investigate the

relationship between MR lesion images and the tissue response. Although other ablation

studies in liver and brain support our findings (8-10), an effort to investigate various

clinically relevant tissue types, should uncover any associated effects on the ability of

MR temperature and thermal lesion images to predict the tissue response. We performed

thermal ablations in normal tissue without examining if a tumor affects MR temperature

and thermal lesion measurements. For our cell death model that accounts for the cellular

response to temperature history, we expect similar results in tumor and other tissue types.

As the temperature increases during a thermal ablation, tissue damage occurs due to

excessive physical forces between molecules which can lead to various destructive cell

processes such as rupturing of cell membranes, denaturation of proteins, and damage of ion channels (11-15). The tissue’s vulnerability to thermal damage depends on molecular structure and strength of chemical bonds but not on cell size or shape. Therefore, the susceptibility to heat damage should be quite similar across different tissue types.

135

Furthermore, studies with normal tissue are most likely sufficient because often it is

clinically desirable to treat the tumor plus a margin of normal tissue. To confirm this,

experiments could be performed on a tumor model, such as the rabbit VX2 tumor model.

6.2.3 Alternative Ablation Techniques and Drug Delivery Therapy

Interventional image-guided ablation therapies using thermal energy sources other than

RF such as laser, microwave, high intensity focused ultrasound, and cryogenics have

received much attention. The spatial gradient of temperature at the margin could be

compared for different thermal ablation techniques. An ablation technique that produces a

steeper temperature gradient is expected to destroy more tumor cells and fewer normal

cells. In addition, the tissue response from various ablation modalities could be examined

using histology images to compare with the tissue response from RF ablation.

We could investigate the use of MRI to guide drug delivery for various

pathologies. A drug, tagged with a MR molecular imaging probe such as caged

gadolinium, would be injected and continuous monitored with MRI to evaluate its spatial-temporal distribution. A high intensity focused ultrasound (HIFU) system would be used to regulate the cellular uptake of the drug via mechanical agitation of the cell membrane to increase its permeability. The HIFU system would be located in the table in the MR system and coupled to the patient via a gel reservoir. An effective delivery of the drug would be characterized by a high intensity MR signal inside a pre-planned target region surrounded by a low MR signal. Computer algorithms would be developed to automatically measure drug dose of each MR voxel from the sequence of MR images.

Regions with an insufficient dose could be identified in the MR images and targeted with

136

HIFU to increase the uptake of the drug. In addition, MRI-guided HIFU thermal ablations could be performed when pathological cells are unresponsive to the drug therapy, or drug

delivery is insufficient.

6.2.4 Novel Registration Methods

For our investigation, our registration method utilized fiducial needles placed near the

region of interest. To expand upon this technique, one could develop an injectable,

biocompatible point fiducial that can be seen in various medical images, such as MR,

ultrasound, CT, positron emission tomography (PET), and localized in macroscopic

tissue and histology. One possibility is to mix matrigel, an extracellular matrix material,

with an imaging specific contrast agent such as gadopentate dimeglumine, iron particles,

and India ink for MR, CT, and tissue, respectively. Multiple point fiducials could be

localized in 3D medical and tissue images, and registered using our current software.

For small animals, such as rats and mice, used in molecular imaging studies, a

methodology that utilizes only 3D anatomical information could be investigated to align

medical images with the tissue response. To correlate MR images to histology, MR and

computed tomography (CT) image volumes of the entire animal could be acquired. After

the intact animal was frozen, the tissue could be sliced and photographed using a

cryotome. The animal’s entire skeleton could then be segmented in the CT and tissue

image volumes, and aligned in 3D using various surface registration techniques. Finally,

the MR image volume could be aligned to the CT volume using voxel-based registration

methods, such as mutual information.

137

6.2.5 Molecular Imaging for Thermal Ablation Guidance

Recently, reports have appeared about novel reversible contrast agents that are activated when coupled to a targeted molecule. Potentially, such molecular imaging probes could be used to monitor thermal ablation therapy. The imaging agents could be targeted to critical molecules in the pathologic cells. As the cell is heated during therapy, morphological changes of the targeted molecules due to denaturation would cause the target and imaging probe to decouple, thereby deactivating the reversible contrast agent.

With probes characterized by short deactivation times, MR signal changes could be monitored for near real-time image guidance.

In conclusion, the continued development of thermal ablation and drug delivery studies can help improve understanding of the relationship between 3D medical scanner images and tissue response. These investigations can lead to changes in the application of

RF, laser, or other energy sources to enhance image-guided ablation therapies, and the development of novel image-guided drug delivery strategies. Furthermore, many clinical and animal model studies seeking to characterize tissue response to ablation or drug delivery would benefit from this research.

138

WORKS CITED

1. Breen, M. S., Wilson, D. L., Lazebnik, R. S., and Lewin, J. S. Three Dimensional Comparison of Interventional MR Radiofrequency Ablation Images With Tissue Response. Medical Image Computing and Computer-Assisted Intervention, Lecture Notes in Computer Science 2003;2879:24-31.

2. Lazebnik, R. S., Weinberg, B. D., Breen, M. S., Lewin, J. S., and Wilson, D. L. Sub-Acute Changes in Lesion Conspicuity and Geometry Following MR-Guided Radiofrequency Ablation. J Magn Reson.Imaging 2003;18(3):353-9.

3. Lazebnik, R. S., Breen, M. S., Fitzmaurice, M., Nour, S. G., Lewin, J. S., and Wilson, D. L. Radio-Frequency-Induced Thermal Lesions: Subacute Magnetic Resonance Appearance and Histological Correlation. J.Magn Reson.Imaging 2003;18(4):487-95.

4. Bederson, J. B., Pitts, L. H., Germano, S. M., Nishimura, M. C., Davis, R. L., and Bartkowski, H. M. Evaluation of 2,3,5-Triphenyltetrazolium Chloride As a Stain for Detection and Quantification of Experimental Cerebral Infarction in Rats. Stroke 1986;17(6):1304-8.

5. Suzuki, M., Zhou, H., Hei, T. K., Tsuruoka, C., and Fujitaka, K. Induction of a Bystander Chromosomal Damage of He-Ion Microbeams in Mammalian Cells. Biol.Sci.Space 2003;17(3):251-2.

6. Shao, C., Stewart, V., Folkard, M., Michael, B. D., and Prise, K. M. Nitric Oxide- Mediated Signaling in the Bystander Response of Individually Targeted Glioma Cells. Cancer Res. 12-1-2003;63(23):8437-42.

7. Gerashchenko, B. I. and Howell, R. W. Cell Proximity Is a Prerequisite for the Proliferative Response of Bystander Cells Co-Cultured With Cells Irradiated With Gamma-Rays. Cytometry 2003;56A(2):71-80.

8. Merkle, E. M., Boll, D. T., Boaz, T., Duerk, J. L., Chung, Y. C., Jacobs, G. H., Varnes, M. E., and Lewin, J. S. MRI-Guided Radiofrequency Thermal Ablation of Implanted VX2 Liver Tumors in a Rabbit Model: Demonstration of Feasibility at 0.2 T. Magn Reson.Med. 1999;42(1):141-9.

9. Morrison, P. R., Jolesz, F. A., Charous, D., Mulkern, R. V., Hushek, S. G., Margolis, R., and Fried, M. P. MRI of Laser-Induced Interstitial Thermal Injury in an in Vivo Animal Liver Model With Histologic Correlation. J.Magn Reson.Imaging 1998;8(1):57-63.

10. Chen, L., Bouley, D. M., Harris, B. T., and Butts, K. MRI Study of Immediate Cell Viability in Focused Ultrasound Lesions in the Rabbit Brain. J.Magn Reson.Imaging 2001;13(1):23-30.

139

11. Bischof, J. C., Padanilam, J., Holmes, W. H., Ezzell, R. M., Lee, R. C., Tompkins, R. G., Yarmush, M. L., and Toner, M. Dynamics of Cell Membrane Permeability Changes at Supraphysiological Temperatures. Biophys.J 1995;68(6):2608-14.

12. Coss, R. A., Dewey, W. C., and Bamburg, J. R. Effects of Hyperthermia on Dividing Chinese Hamster Ovary Cells and on Microtubules in Vitro. Cancer Res. 1982;42(3):1059-71.

13. Borrelli, M. J., Wong, R. S., and Dewey, W. C. A Direct Correlation Between Hyperthermia-Induced Membrane Blebbing and Survival in Synchronous G1 CHO Cells. J Cell Physiol 1986;126(2):181-90.

14. Dewey, W. C. Failla Memorial Lecture. The Search for Critical Cellular Targets Damaged by Heat. Radiat.Res. 1989;120(2):191-204.

15. Coss, R. A., Dewey, W. C., and Bamburg, J. R. Effects of Hyperthermia on Dividing Chinese Hamster Ovary Cells and on Microtubules in Vitro. Cancer Res. 1982;42(3):1059-71.

140

APPENDIX 1

Principles of Magnetic Resonance Imaging

The basis of clinical magnetic resonance imaging (MRI) is the ability to accurately localize the magnetic properties of the nucleus of the atom to create an image. The MR image can be used to examine anatomic and physiologic properties of the patient, and to measure the temperature distribution of the tissue. This appendix reviews the generation of tissue contrast from specific pulse sequences, and the details of MR thermometry.

A1.1 Contrast Mechanisms

A variety of biologically relevant elements may be used to produce MR images. Hydrogen (1H), with the largest magnetic moment and greatest physiologic concentration, is the best element for general clinical use. Other elements such as 23Na, 31P have been used for imaging but are orders of magnitude less sensitive. Therefore, the proton is the principal element used for MR imaging.

An MR signal is produced by applying a uniform static magnetic field and a pulse of radiofrequency (RF) energy. Thermal energy at physiologic temperatures agitates and randomizes the orientation of the magnetic moments of the protons in a tissue sample, and there is no net tissue magnetization. However, a strong magnetic field (Bo) applied to a group of protons will generate an equilibrium magnetization (Mo) in the direction of Bo (longitudinal magnetization), with no magnetization in the perpendicular direction (transverse magnetization).

In addition, the protons will precess (spin) about their axis, parallel to Bo, with an angular frequency that is proportional to the applied magnetic field strength, according to the Larmor equation. The application of pulse of RF energy (B1) at the precessional frequency of the protons

141 causes displacement of the tissue magnetization from the equilibrium direction. As the magnetization returns to equilibrium conditions, MR signals are emitted.

The MR signals produced as the tissue magnetization returns to equilibrium are proportional to the number of excited protons in the sample, with a rate that depends on the T2 and T1 relaxation characteristics of the tissue. For T2 relaxation, a 90-degree RF pulse, which rotates the magnetization direction by 90-degrees, produces phase coherence of the individual protons of a tissue volume and generates the maximum possible transverse magnetization (Mxy). As Mxy rotates at the Larmor frequency, the receiver antenna coil is induced to produce a damped sinusoidal signal known as the free induction decay (FID) signal. Exponential relaxation decay,

T2, represents the intrinsic spin-spin interactions that cause loss of phase coherence due to intrinsic magnetic properties of the sample. The elapsed time between the peak transverse signal and 37% of the peak level (1/e) is the T2 decay constant. Mathematically, this relationship is expressed as follows:

−tT/2 M(xy t)=M0e (A1.1)

T2 decay mechanisms are determined by the molecular structure of the tissue sample.

Magnetic inhomogeneities intrinsic to the structure of the sample cause a spin-spin interaction, where the individual spins precess at different frequencies due to slight changes in the intrinsic local magnetic field strength. Small mobile molecules in liquids have long T2 because rapid molecular motion reduces intrinsic magnetic inhomogeneities. As molecular size increases, limited molecular motion causes larger intrinsic magnetic field variations and the T2 decay is more rapid (short T2).

The loss of transverse magnetization occurs rapidly relative to the return of the longitudinal magnetization (Mz) to equilibrium (Mz = Mo). The T1 relaxation describes the exponential

142 regrowth of Mz, which depends on the characteristics of the spin interaction with the lattice

(molecular arrangement and structure of the tissue sample). The T1 relaxation constant is the time needed to recover 63% of Mz, after a 90-degree pulse (when Mz = 0). The recovery of Mz versus time after the 90-degree RF pulse is expressed mathematically as follows:

−tT/1 M(z0t)=M(1− e ) (A1.2)

The T1 relaxation time depends on the dissipation rate of the absorbed energy into the surrounding molecular tissue (lattice). The energy transfer is most efficient when the precessional (Larmor) frequency of the excited protons overlaps with the vibrational frequencies of the molecular lattice. Medium size molecules, such as proteins and viscous fluids, produce significant vibrations at frequencies near the Larmor frequency. Hence, these molecules have a short T1 relaxation time. However, small and large molecules have small vibrations near the

Larmor frequency, thereby creating long T1 relaxations. The commonly used contrast agent, gadolinium-diethylenetriaminepentaacetic acid (Gd-DTPA), dramatically decreases the T1 relaxation time since free protons become bound and a hydration layer is formed, thus providing a spin-lattice energy sink and a rapid return to equilibrium.

The exquisite contrast sensitivity of MR images can be achieved by emphasizing the differences among T1 and T2 relaxation time constants, and proton density of the tissues. The pulse sequence design, which consists of the timing, order, polarity, and repetition frequency of the applied RF pulses and magnetic field gradients, makes the detected signals dependent on either proton density, T1 or T2 relaxation times. Two common MR pulse sequences are spin echo (SE) and gradient recalled echo (GRE). Using these sequences combined with localization methods, which spatially encode the signal, allow contrast-weighted MR images to be obtained.

143 For a spin echo sequence, the magnetized protons in a sample are excited with an initial 90-

degree RF pulse to produce FID, followed by a 180-degree RF pulse applied at time TE/2 to produce an echo at time TE. The signal is acquired during the creation and decay of the echo.

After a delay time allowing recovery of Mz (the end of the TR period), the sequence is repeated many times to acquire the information needed to create an image. The repetition time, TR, is the time between 90-degree excitation pulses. The timing between the RF pulses allows for adjustment of tissue contrast. The signal produced can be mathematically expressed as:

−−TR /1T TE /T 2 S ∝−ρHf(v)(1 e)e (A1.3)

where ρH is the proton density, f(v) is the signal from fluid flow, T1 and T2 are physical

properties of tissue, and TR and TE are pulse sequence parameters.

A T1-weighted SE sequence produces contrast mainly from the T1 characteristics of tissues

by de-emphasizing T2 and proton density contributions. To minimize T2 contrast, a short TE

(typically 5-30 msec) is used. To lessen the proton density contrast relative to the T1 contrast, an

optimal TR must be selected. Typically, a TR on the order of the T1 values of the tissues of

interest (400-600 msec) will be used to emphasize the T1 contrast as compared to the proton

density contrast. This will obviously require some a priori knowledge of tissue properties.

Proton density weighting depends on differences in the number of magnetized protons per

volume of tissue. Tissues with greater spin density have a larger longitudinal magnetization.

Hydrogenous tissues such as fat have a high spin density as compared with soft tissue, which

contain much protein. To minimize T1 differences in the tissues, a relatively long TR is used to

allow significant longitudinal magnetization recovery to equilibrium. The signal amplitudes in

the FID are preserved with a short TE, which minimizes the influences of T2 differences.

144 A T2-weighted SE sequence emphasizes differences in T2 characteristics of tissues. This is

accomplished by using a relatively long TR (1,500-3,500 msec) to reduce T1 effects. To

emphasize the T2 contrast relative to the proton density contrast, an optimal TE is needed. In

general, a TE on the order of the T2 values of the tissues of interest (typically 60-150 msec) will

enhance the T2 differences as compared to the proton density differences.

The gradient recalled echo (GRE) technique uses a magnetic field gradient, instead of the

180-degree RF pulse, to produce an echo. Dephasing the transverse magnetization spins with an

applied gradient of one polarity and rephrasing the spins with the gradient reversed in polarity

produces a gradient recalled echo. Tissue contrast in GRE pulse sequences depends on TR, TE,

and the flip angle. An advantage of GRE sequences is the short TR, which permits fast imaging

as compared to SE sequences.

A1.2 MR Thermometry

Various MR-measured parameters can be used to map temperature changes. Temperature MRI

based on proton resonance frequency (PRF) of tissue water, which provides excellent linearity,

near-independence with respect to tissue type, and good temperature sensitivity, is the preferred

thermometry method for many applications. Other MRI temperature methods are based on the

T1 relaxation time and diffusion coefficient but nonlinear effects are observed. Therefore, the

PRF-based method is generally used for MR thermometry.

The PRF temperature dependency is based on changes to intermolecular forces and

hydrogen bond formation between water molecules. The local magnetic field Bnuc(r) as observed by the spins is a function of the main magnetic field B0 and the chemical shift σ(T(r)):

Bnuc(T) = [1+σ (T)]B0 (A1.4)

145 The chemical shift field (in ppm) is the sum of the temperature independent contributions, for

example those from the B0 field inhomogeneities, σ0, and a temperature-dependent contribution

σT(T):

σ (T)=σσ0T + (T) = σ0 + αT (A1.5)

where α, is the temperature-dependent water chemical shift in ppm °C-1 (approximately

0.01 ppm • °C-1). The chemical shift field can be calculated from the phase information in

gradient-echo (GE) images:

Φ(T)=γσ (T)TE0B (A1.6) where Φ is the image phase, γ is the gyromagnetic ratio of the observed nucleus

(42.58 106 Hz T-1 for protons), and TE is the echo time. To measure temperature-dependent

changes in chemical shift, σ0 must be removed. This can be accomplished by subtraction of the

magnetic field distribution measured at a given reference temperature from the field distribution

measured at temperature T. This leads to: Φ(T)-Φ(T) ∆T = T-Tref = (A1.7) αγTBE0

The PRF method uses GE imaging techniques to measure the phase change resulting

from the temperature-dependent change in resonant frequency. Spin-echo imaging sequences

cannot be used since the temperature-induced phase contribution will be refocused.

The phase difference between reference data and data acquired during heating increases

linearly with the TE of the experiment. However, the image signal to noise ratio (SNR) decreases

exponentially as a function of TE. Therefore, the optimal TE for the phase SNR, and thus for the

temperature sensitivity of the sequence, is equal to the T2* value of the tissue. This can be easily

derived from the expression:

146 -TTE/ 2* SNRTE = cT e (A1.8)

Where SNRT is the SNR of the temperature map, and c is a constant that depends on spin density,

flip angle, TR, field strength, and RF coil.

An important advantage of the PRF method is its near-independence of the tissue

composition . However, the presence of lipids is a potential source of artifacts since the PRFs of

lipid hydrogens are independent of temperature. Using a fat suppression imaging technique,

these artifacts can be minimized.

147 BIBLIOGRAPHY

Aschoff, A. J., Rafie, N., Jesberger, J. A., Duerk, J. L., and Lewin, J. S. Thermal Lesion Conspicuity Following Interstitial Radiofrequency Thermal Tumor Ablation in Humans: a Comparison of STIR, Turbo Spin-Echo T2-Weighted, and Contrast-Enhanced T1- Weighted MR Images at 0.2 T. J.Magn Reson.Imaging 2000;12(4):584-9.

Baroldi, G. Myocardial cell death, including ischemic heart disease and its complications. Silver, M. D., Gotlied, A. I., and Schoen, F. J. Cardiovascular pathology. 3 ed.Churchill Livingstone, NY; 2001. pp.202-6.

Bederson, J. B., Pitts, L. H., Germano, S. M., Nishimura, M. C., Davis, R. L., and Bartkowski, H. M. Evaluation of 2,3,5-Triphenyltetrazolium Chloride As a Stain for Detection and Quantification of Experimental Cerebral Infarction in Rats. Stroke 1986;17(6):1304-8.

Besl, P. J. and McKay, H. D. A Method for Registration of 3-D Shapes. IEEE Transactions on Pattern Analysis and Machine Intelligence 1992;14(2):239-56.

Bischof, J. C., Padanilam, J., Holmes, W. H., Ezzell, R. M., Lee, R. C., Tompkins, R. G., Yarmush, M. L., and Toner, M. Dynamics of Cell Membrane Permeability Changes at Supraphysiological Temperatures. Biophys.J 1995;68(6):2608-14.

Boaz, T. L., Lewin, J. S., Chung, Y. C., Duerk, J. L., Clampitt, M. E., and Haaga, J. R. MR Monitoring of MR-Guided Radiofrequency Thermal Ablation of Normal Liver in an Animal Model. J.Magn Reson.Imaging 1998;8(1):64-9.

Bookstein, F. L. Thin-Plate Splines and the Decomposition of Deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence 1989;11(6):567-85.

Borrelli, M. J., Wong, R. S., and Dewey, W. C. A Direct Correlation Between Hyperthermia-Induced Membrane Blebbing and Survival in Synchronous G1 CHO Cells. J Cell Physiol 1986;126(2):181-90.

Breen, M. S., Lancaster, T. L., Lazebnik, R. S., Nour, S. G., Lewin, J. S., and Wilson, D. L. Three-Dimensional Method for Comparing in Vivo Interventional MR Images of Thermally Ablated Tissue With Tissue Response. J Magn Reson.Imaging 2003;18(1):90- 102.

Breen, M. S., Wilson, D. L., Lazebnik, R. S., and Lewin, J. S. Three Dimensional Comparison of Interventional MR Radiofrequency Ablation Images With Tissue Response. Medical Image Computing and Computer-Assisted Intervention, Lecture Notes in Computer Science 2003;2879:24-31.

Breen, M. S., Lazebnik, R. S., Fitzmaurice, M., Nour, S. G., Lewin, J. S., and Wilson, D. L. Radio-Frequency Thermal Ablation: Correlation of Hyperacute MR Lesion Images With Tissue Response. J Magn Reson.Imaging 2004;Submitted.

148 Bromer, R. H., Mitchell, J. B., and Soares, N. Response of Human Hematopoietic Precursor Cells (CFUc) to Hyperthermia and Radiation. Cancer Res. 1982;42(4):1261-5.

Bruneval, P., Mesnildrey, P., and Camilleri, J. P. Nd-YAG Laser-Induced Injury in Dog Myocardium: Optical and Ultrastructural Study of Early Lesions. Eur.Heart J 1987;8(7):785-92.

Chalana, V. and Kim, Y. A Methodology for Evaluation of Boundary Detection Algorithms on Medical Images. IEEE Transactions on Medical Imaging 2002;16(5):642- 52.

Chen, L., Bouley, D. M., Harris, B. T., and Butts, K. MRI Study of Immediate Cell Viability in Focused Ultrasound Lesions in the Rabbit Brain. J.Magn Reson.Imaging 2001;13(1):23-30.

Chen, L., Wansapura, J. P., Heit, G., and Butts, K. Study of Laser Ablation in the in Vivo Rabbit Brain With MR Thermometry. J Magn Reson.Imaging 2002;16(2):147-52.

Cline, H. E., Schenck, J. F., Watkins, R. D., Hynynen, K., and Jolesz, F. A. Magnetic Resonance-Guided Thermal Surgery. Magn Reson.Med. 1993;30(1):98-106.

Cline, H. E., Hynynen, K., Watkins, R. D., Adams, W. J., Schenck, J. F., Ettinger, R. H., Freund, W. R., Vetro, J. P., and Jolesz, F. A. Focused US System for MR Imaging- Guided Tumor Ablation. Radiology 1995;194(3):731-7.

Cosman, E. R., Nashold, B. S., and Ovelman-Levitt, J. Theoretical Aspects of Radiofrequency Lesions in the Dorsal Root Entry Zone. Neurosurgery 1984;15(6):945- 50.

Coss, R. A., Dewey, W. C., and Bamburg, J. R. Effects of Hyperthermia on Dividing Chinese Hamster Ovary Cells and on Microtubules in Vitro. Cancer Res. 1982;42(3):1059-71.

Cutts, A. Shrinkage of Muscle Fibres During the Fixation of Cadaveric Tissue. J.Anat. 1988;160:75-8.

Davis, M. H., Khotanzad, A., Flamig, D. P., and Harris, B. T. A Physics-Based Coordinate Transformation for 3-D Image Matching. IEEE Trans.Med.Imaging 1997;16(3):317-28.

Dewey, W. C. Failla Memorial Lecture. The Search for Critical Cellular Targets Damaged by Heat. Radiat.Res. 1989;120(2):191-204.

Falcao, A. X., Udupa, J. K., Samarasekera, S., Sharma, S., Hirsch, B. E., and Lotufo, D. A. User-Steered Image Segmentation Paradigms: Live Wire and Live Lane. Graphical Models and Image Processing 1998;60(4):233-60.

149 Farahani, K., Mischel, P. S., Black, K. L., De Salles, A. A., Anzai, Y., and Lufkin, R. B. Hyperacute Thermal Lesions: MR Imaging Evaluation of Development in the Brain. Radiology 1995;196(2):517-20.

Farrell, M. A., Charboneau, W. J., DiMarco, D. S., Chow, G. K., Zincke, H., Callstrom, M. R., Lewis, B. D., Lee, R. A., and Reading, C. C. Imaging-Guided Radiofrequency Ablation of Solid Renal Tumors. AJR Am.J Roentgenol. 2003;180(6):1509-13.

Gazelle, G. S., Goldberg, S. N., Solbiati, L., and Livraghi, T. Tumor Ablation With Radio-Frequency Energy. Radiology 2000;217(3):633-46.

Gerashchenko, B. I. and Howell, R. W. Cell Proximity Is a Prerequisite for the Proliferative Response of Bystander Cells Co-Cultured With Cells Irradiated With Gamma-Rays. Cytometry 2003;56A(2):71-80.

Gershfeld, N. L. and Murayama, M. Thermal Instability of Red Blood Cell Membrane Bilayers: Temperature Dependence of Hemolysis. J Membr.Biol. 1988;101(1):67-72.

Ghanem, I., Collet, L. M., Kharrat, K., Samaha, E., Deramon, H., Mertl, P., and Dagher, F. Percutaneous Radiofrequency Coagulation of Osteoid Osteoma in Children and Adolescents. J Pediatr.Orthop.B 2003;12(4):244-52.

Goldberg, S. N., Gazelle, G. S., Halpern, E. F., Rittman, W. J., Mueller, P. R., and Rosenthal, D. I. Radiofrequency Tissue Ablation: Importance of Local Temperature Along the Electrode Tip Exposure in Determining Lesion Shape and Size. Acad.Radiol. 1996;3(3):212-8.

Goldberg, S. N.; Livraghi, T.; Solbiati, L.; Gazelle, G. S. In situ ablation of focal hepatic neoplasms. Gazelle, G. S., Saini S., and Mueller, P. R. Hepatobiliary and pancreatic radiology: imaging and intervention. New York: Thiema Medical Pub; 1997.

Goldberg, S. N., Gazelle, G. S., and Mueller, P. R. Thermal Ablation Therapy for Focal Malignancy: a Unified Approach to Underlying Principles, Techniques, and Diagnostic Imaging Guidance. AJR Am.J Roentgenol. 2000;174(2):323-31.

Graham, S. J., Stanisz, G. J., Kecojevic, A., Bronskill, M. J., and Henkelman, R. M. Analysis of Changes in MR Properties of Tissues After Heat Treatment. Magn Reson.Med. 1999;42(6):1061-71.

Graham, S. J., Chen, L., Leitch, M., Peters, R. D., Bronskill, M. J., Foster, F. S., Henkelman, R. M., and Plewes, D. B. Quantifying Tissue Damage Due to Focused Ultrasound Heating Observed by MRI. Magn Reson.Med. 1999;41(2):321-8.

Hall-Craggs, M. A. and Vaidya, J. S. Minimally Invasive Therapy for the Treatment of Breast Tumours. Eur.J Radiol 2002;42(1):52-7.

150 Hayes, L. J., Diller, K. R., Pearce, J. A., Schick, M. R., and Colvin, D. P. Prediction of Transient Temperature Fields and Cumulative Tissue Destruction for Radio Frequency Heating of a Tumor. Med.Phys. 1985;12(6):684-92.

Hazle, J. D., Stafford, R. J., and Price, R. E. Magnetic Resonance Imaging-Guided Focused Ultrasound Thermal Therapy in Experimental Animal Models: Correlation of Ablation Volumes With Pathology in Rabbit Muscle and VX2 Tumors. J Magn Reson.Imaging 2002;15(2):185-94.

Izzo, F. Other Thermal Ablation Techniques: Microwave and Interstitial Laser Ablation of Liver Tumors. Ann.Surg.Oncol. 2003;10(5):491-7.

Jacobs, M. A., Windham, J. P., Soltanian-Zadeh, H, Peck, D. J., and Knight, R. A. Registration and Warping of Magnetic Resonance Images to Histological Sections. Med.Phys. 1999;26(8):1568-78.

Johnson, P. C. and Saidel, G. M. Thermal Model for Fast Simulation During Magnetic Resonance Imaging Guidance of Radio Frequency Tumor Ablation. Ann.Biomed.Eng 2002;30(9):1152-61.

Jolesz, F. A., Bleier, A. R., Jakab, P., Ruenzel, P. W., Huttl, K., and Jako, G. J. MR Imaging of Laser-Tissue Interactions. Radiology 1988;168(1):249-53.

Jolesz, F. A. and Shtern, F. The Operating Room of the Future. Report of the National Cancer Institute Workshop, "Imaging-Guided Stereotactic Tumor Diagnosis and Treatment". Invest Radiol 1992;27(4):326-8.

Jolesz, F. A. and Blumenfeld, S. M. Interventional Use of Magnetic Resonance Imaging. Magn Reson.Q. 1994;10(2):85-96.

Kahn, T., Bettag, M., Ulrich, F., Schwarzmaier, H. J., Schober, R., Furst, G., and Modder, U. MRI-Guided Laser-Induced Interstitial Thermotherapy of Cerebral Neoplasms. J Comput.Assist.Tomogr. 1994;18(4):519-32.

Kahn, T., Harth, T., Kiwit, J. C., Schwarzmaier, H. J., Wald, C., and Modder, U. In Vivo MRI Thermometry Using a Phase-Sensitive Sequence: Preliminary Experience During MRI-Guided Laser-Induced Interstitial Thermotherapy of Brain Tumors. J Magn Reson.Imaging 1998;8(1):160-4.

Lancaster, T. L and Wilson, D. L. Correcting Spatial Distortion of Histological Images. Annals of Biomedical Engineering 2003;(In Press).

Larson, T. R., Bostwick, D. G., and Corica, A. Temperature-Correlated Histopathologic Changes Following Microwave Thermoablation of Obstructive Tissue in Patients With Benign Prostatic Hyperplasia. Urology 1996;47(4):463-9.

151 Lazebnik, R. S., Weinberg, B. D., Breen, M. S., Lewin, J. S., and Wilson, D. L. Three- Dimensional Model of Lesion Geometry for Evaluation of MR-Guided Thermal Ablation Therapy. Acad.Radiol 2002;9(10):1128-38.

Lazebnik, R. S., Lancaster, T. L., Breen, M. S., Lewin, J. S., and Wilson, D. L. Volume Registration Using Needle Paths and Point Landmarks for Evaluation of Interventional MRI Treatments. IEEE Trans.Med.Imaging 2003;22(5):653-60.

Lazebnik, R. S., Breen, M. S., Fitzmaurice, M., Nour, S. G., Lewin, J. S., and Wilson, D. L. Radio-Frequency-Induced Thermal Lesions: Subacute Magnetic Resonance Appearance and Histological Correlation. J Magn Reson.Imaging 2003;18(4):487-95.

Lazebnik, R. S., Weinberg, B. D., Breen, M. S., Lewin, J. S., and Wilson, D. L. Sub- Acute Changes in Lesion Conspicuity and Geometry Following MR-Guided Radiofrequency Ablation. J Magn Reson.Imaging 2003;18(3):353-9.

Lee, J. D., Lee, J. M., Kim, S. W., Kim, C. S., and Mun, W. S. MR Imaging- Histopathologic Correlation of Radiofrequency Thermal Ablation Lesion in a Rabbit Liver Model: Observation During Acute and Chronic Stages. Korean J Radiol 2001;2:151-8.

Lee, R. C. and Astumian, R. D. The Physicochemical Basis for Thermal and Non- Thermal 'Burn' Injuries. Burns 1996;22(7):509-19.

Lee, Z., Sodee, D. B, Faulhaber, P. F., Lancaster, T. L, MacLennan, G. T., and Wilson, D. L. Comparison of SPECT and PET Imaging for Prostate Cancer With Histological Correlation. Journal of Nuclear Medicine 2001;42:294 (Abstract).

Lepock, J. R. Involvement of Membranes in Cellular Responses to Hyperthermia. Radiat.Res. 1982;92(3):433-8.

Lewin, J. S., Duerk, J. L., Jain, V. R., Petersilge, C. A., Chao, C. P., and Haaga, J. R. Needle Localization in MR-Guided Biopsy and Aspiration: Effects of Field Strength, Sequence Design, and Magnetic Field Orientation. AJR Am.J.Roentgenol. 1996;166(6):1337-45.

Lewin, J. S., Connell, C. F., Duerk, J. L., Chung, Y. C., Clampitt, M. E., Spisak, J., Gazelle, G. S., and Haaga, J. R. Interactive MRI-Guided Radiofrequency Interstitial Thermal Ablation of Abdominal Tumors: Clinical Trial for Evaluation of Safety and Feasibility. J.Magn Reson.Imaging 1998;8(1):40-7.

Leyko, W. and Bartosz, G. Membrane Effects of Ionizing Radiation and Hyperthermia. Int.J Radiat.Biol.Relat Stud.Phys.Chem.Med. 1986;49(5):743-70.

152 Likar, B. and Pernus, F. Registration of Serial Transverse Sections of Muscle Fibers. Cytometry 1999;37(2):93-106.

Liu, H., Hall, W. A., Martin, A. J., and Truwit, C. L. Biopsy Needle Tip Artifact in MR- Guided Neurosurgery. J.Magn Reson.Imaging 2001;13(1):16-22.

Livraghi, T., Solbiati, L., Meloni, F., Ierace, T., Goldberg, S. N., and Gazelle, G. S. Percutaneous Radiofrequency Ablation of Liver Metastases in Potential Candidates for Resection: the "Test-of-Time Approach". Cancer 6-15-2003;97(12):3027-35.

Lo, H. W., Tsai, Y. J., Chen, P. H., Chen, H. Y., Ker, C. G., and Juan, C. C. Radiofrequency Ablation for Treatment of Hepatocellular Carcinoma With Cirrhosis. Hepatogastroenterology 2003;50(51):645-50.

Mack, M. G. and Vogl, T. J. MR-Guided Ablation of Head and Neck Tumors. Magn Reson.Imaging Clin.N.Am. 2002;10(4):707-13, vi.

Maintz, J. B. and Viergever, M. A. A Survey of Medical Image Registration. Med.Image Anal. 1998;2(1):1-36.

McDannold, Hynynen, K., Wolf, D., Wolf, G., and Jolesz, F. MRI Evaluation of Thermal Ablation of Tumors With Focused Ultrasound. J Magn Reson.Imaging 1998;8(1):91-100.

McDannold, N. J., King, R. L., Jolesz, F. A., and Hynynen, K. H. Usefulness of MR Imaging-Derived Thermometry and Dosimetry in Determining the Threshold for Tissue Damage Induced by Thermal Surgery in Rabbits. Radiology 2000;216(2):517-23.

Mega, M. S, Chen, S. S, Thompson, P. M., Woods, R. P., Karaca, T. J., Tiwari, A., Vinters, H. V., Small, G. W, and Toga, A. W. Mapping Histology to Metabolism: Coregistration of Stained Whole-Brain Sections to Premortem PET in Alzheimer's Disease. Neuroimage 1997;5(2):147-53.

Merkle, E. M., Haaga, J. R., Duerk, J. L., Jacobs, G. H., Brambs, H. J., and Lewin, J. S. MR Imaging-Guided Radio-Frequency Thermal Ablation in the Pancreas in a Porcine Model With a Modified Clinical C-Arm System. Radiology 1999;213(2):461-7.

Merkle, E. M., Boll, D. T., Boaz, T., Duerk, J. L., Chung, Y. C., Jacobs, G. H., Varnes, M. E., and Lewin, J. S. MRI-Guided Radiofrequency Thermal Ablation of Implanted VX2 Liver Tumors in a Rabbit Model: Demonstration of Feasibility at 0.2 T. Magn Reson.Med. 1999;42(1):141-9.

Merkle, E. M., Shonk, J. R., Zheng, L., Duerk, J. L., and Lewin, J. S. MR Imaging- Guided Radiofrequency Thermal Ablation in the Porcine Brain at 0.2 T. Eur.Radiol. 2001;11:884-92.

Metz, C. E. Basic Principles of ROC Analysis. Semin.Nucl.Med. 1978;8(4):283-98.

153 Meyer, C. R., Leichtman, G. S., Brunberg, J. A., Wahl, R. L., and Quint, L. E. Simultaneous Usage of Homologous Points, Lines, and Planes for Optimal, 3-D, Linear Registration of Multimodality Imaging Data. IEEE Transactions on Medical Imaging 1995;14(1):1-11.

Mirza, A. N., Fornage, B. D., Sneige, N., Kuerer, H. M., Newman, L. A., Ames, F. C., and Singletary, S. E. Radiofrequency Ablation of Solid Tumors. Cancer J 2001;7(2):95- 102.

Morrison, P. R., Jolesz, F. A., Charous, D., Mulkern, R. V., Hushek, S. G., Margolis, R., and Fried, M. P. MRI of Laser-Induced Interstitial Thermal Injury in an in Vivo Animal Liver Model With Histologic Correlation. J.Magn Reson.Imaging 1998;8(1):57-63.

Muralidharan, V., Malcontenti-Wilson, C., and Christophi, C. Interstitial Laser Hyperthermia for Colorectal Liver Metastases: the Effect of Thermal Sensitization and the Use of a Cylindrical Diffuser Tip on Tumor Necrosis. J Clin.Laser Med.Surg. 2002;20(4):189-96.

Nelder, J. A. and Mead, R. A Simplex Method for Function Minimization. Computer Journal 1965;7:308-13.

Nour, S. G., Aschoff, A. J., Mitchell, I. C., Emancipator, S. N., Duerk, J. L., and Lewin, J. S. MR Imaging-Guided Radio-Frequency Thermal Ablation of the Lumbar Vertebrae in Porcine Models. Radiology 2002;224(2):452-62.

Peters, R. D., Chan, E., Trachtenberg, J., Jothy, S., Kapusta, L., Kucharczyk, W., and Henkelman, R. M. Magnetic Resonance Thermometry for Predicting Thermal Damage: an Application of Interstitial Laser Coagulation in an in Vivo Canine Prostate Model. Magn Reson.Med. 2000;44(6):873-83.

Puccini, S., Bar, N. K., Bublat, M., Kahn, T., and Busse, H. Simulations of Thermal Tissue Coagulation and Their Value for the Planning and Monitoring of Laser-Induced Interstitial Thermotherapy (LITT). Magn Reson.Med. 2003;49(2):351-62.

Roti Roti, J. L. and Winward, R. T. Factors Affecting the Heat-Induced Increase in Protein Content of Chromatin. Radiat.Res. 1980;81(1):138-44.

Salem, K. A., Szymanski-Exner, A., Lazebnik, R. S, Breen, M. S., Gao, J., and Wilson, D. L. X-Ray Computed Tomography Methods for in Vivo Evaluation of Local Drug Release Systems. IEEE Transactions on Medical Imaging 2002;(In Press).

Sapareto, S. A., Hopwood, L. E., Dewey, W. C., Raju, M. R., and Gray, J. W. Effects of Hyperthermia on Survival and Progression of Chinese Hamster Ovary Cells. Cancer Res. 1978;38(2):393-400.

Sapareto, S. A. and Dewey, W. C. Thermal Dose Determination in Cancer Therapy. Int.J Radiat.Oncol.Biol.Phys. 1984;10(6):787-800.

154 Schenck, J. F., Jolesz, F. A., Roemer, P. B., Cline, H. E., Lorensen, W. E., Kikinis, R., Silverman, S. G., Hardy, C. J., Barber, W. D., Laskaris, E. T., and . Superconducting Open-Configuration MR Imaging System for Image-Guided Therapy. Radiology 1995;195(3):805-14.

Schulze, P. C., Kahn, T., Harth, T., Schwurzmaier, H. J., Schober, R., and Schulze, C. P. Correlation of Neuropathologic Findings and Phase-Based MRI Temperature Maps in Experimental Laser-Induced Interstitial Thermotherapy. J Magn Reson.Imaging 1998;8(1):115-20.

Schwarzmaier, H. J., Yaroslavsky, I. V., Yaroslavsky, A. N., Fiedler, V., Ulrich, F., and Kahn, T. Treatment Planning for MRI-Guided Laser-Induced Interstitial Thermotherapy of Brain Tumors--the Role of Blood Perfusion. J Magn Reson.Imaging 1998;8(1):121-7.

Shao, C., Stewart, V., Folkard, M., Michael, B. D., and Prise, K. M. Nitric Oxide- Mediated Signaling in the Bystander Response of Individually Targeted Glioma Cells. Cancer Res. 12-1-2003;63(23):8437-42.

Sherar, M. D., Moriarty, J. A., Kolios, M. C., Chen, J. C., Peters, R. D., Ang, L. C., Hinks, R. S., Henkelman, R. M., Bronskill, M. J., and Kucharcyk, W. Comparison of Thermal Damage Calculated Using Magnetic Resonance Thermometry, With Magnetic Resonance Imaging Post-Treatment and Histology, After Interstitial Microwave Thermal Therapy of Rabbit Brain. Phys.Med.Biol. 2000;45(12):3563-76.

Spitzer, V. M and Whitlock, D. G. Atlas of the Visible Human Male: Reverse Engineering of the Human Body. Jones and Bartlett 1998.

Stollberger, R., Ascher, P. W., Huber, D., Renhart, W., Radner, H., and Ebner, F. Temperature Monitoring of Interstitial Thermal Tissue Coagulation Using MR Phase Images. J Magn Reson.Imaging 1998;8(1):188-96.

Suzuki, M., Zhou, H., Hei, T. K., Tsuruoka, C., and Fujitaka, K. Induction of a Bystander Chromosomal Damage of He-Ion Microbeams in Mammalian Cells. Biol.Sci.Space 2003;17(3):251-2.

Swets, J. A. ROC Analysis Applied to the Evaluation of Medical Imaging Techniques. Invest Radiol 1979;14(2):109-21.

Thomsen, S., Pearce, J. A., and Cheong, W. F. Changes in Birefringence As Markers of Thermal Damage in Tissues. IEEE Trans.Biomed.Eng 1989;36(12):1174-9.

Thomsen, S. Pathologic Analysis of Photothermal and Photomechanical Effects of Laser- Tissue Interactions. Photochem.Photobiol. 1991;53(6):825-35.

Toga, A. W, Quinn, B., and Ambach, K. A High Resolution Digital Collection of Cryosectioned Whole Brain. J.Cerebal Blood Flow and Metabolism 1993;13(1):820.

155 Toga, A. W., Ambach, K., Quinn, B., Hutchin, M., and Burton, J. S. Postmortem Anatomy From Cryosectioned Whole Human Brain. J.Neurosci.Methods 1994;54(2):239-52.

Toga, A. W., Goldkorn, A., Ambach, K., Chao, K., Quinn, B. C., and Yao, P. Postmortem Cryosectioning As an Anatomic Reference for Human Brain Mapping. Comput.Med.Imaging Graph. 1997;21(2):131-41.

Vogl, T. J., Mack, M. G., Muller, P., Phillip, C., Bottcher, H., Roggan, A., Juergens, M., Deimling, M., Knobber, D., Wust, P., and . Recurrent Nasopharyngeal Tumors: Preliminary Clinical Results With Interventional MR Imaging--Controlled Laser-Induced Thermotherapy. Radiology 1995;196(3):725-33.

Vykhodtseva, N., Sorrentino, V., Jolesz, F. A., Bronson, R. T., and Hynynen, K. MRI Detection of the Thermal Effects of Focused Ultrasound on the Brain. Ultrasound Med.Biol. 2000;26(5):871-80.

Wood, B. J., Ramkaransingh, J. R., Fojo, T., Walther, M. M., and Libutti, S. K. Percutaneous Tumor Ablation With Radiofrequency. Cancer 1-15-2002;94(2):443-51.

156