CT-PET IMAGE FUSION AND PET IMAGE SEGMENTATION FOR RADIATION THERAPY
by
Yiran Zheng
Submitted in partial fulfillment of the requirements
For the degree of Doctor of Philosophy
Dissertation Adviser: Barry W. Wessels, Ph.D.
Department of Biomedical Engineering
CASE WESTERN RESERVE UNIVERSITY
January, 2011
CASE WESTERN RESERVE UNIVERSITY
SCHOOL OF GRADUATE STUDIES
We hereby approve the thesis/dissertation of Yiran Zheng candidate for the Ph.D. degree *.
(signed) Andrew M. Rollins, Ph.D. (chair of the committee)
Xin Yu, Ph.D.
Barry W. Wessels, Ph.D.
Syed F. Akber, Ph.D.
______
______
(date) July 12th, 2010
*We also certify that written approval has been obtained for any proprietary material contained therein.
Dedication
To my parents and my wife
Table of Contents
Table of Contents ...... 1
List of Tables ...... 4
List of Figures ...... 6
Acknowledgements ...... 8
List of Abbreviations ...... 10
Abstract ...... 12
Chapter 1 An Introduction to PET and Radiation Therapy
Treatment Planning ...... 14
1.1 Physics of positron emission tomography (PET) ...... 14
1.1.1 The spatial resolution of PET imaging system ...... 21
1.1.2 18F-FDG - the radiotracer used in PET ...... 24
1.2 Radiation therapy treatment planning ...... 27
1.2.1 Radiation therapy and External Beam Radiation
Therapy ...... 27
1.2.2 Treatment planning for radiation therapy...... 32
1.3 The role of PET in radiation therapy treatment planning
and the significance of this research ...... 35
1.4 Overview of dissertation organization ...... 38
Chapter 2 A Machine Based CT-PET Image Fusion System ...... 41
- 1 -
2.1 Abstract ...... 41
2.2 Introduction ...... 43
2.3 Methods and materials ...... 46
2.3.1 Patients ...... 46
2.3.2 Image Acquisition ...... 46
2.3.3 Image Fusion ...... 49
2.3.4 Data Analysis ...... 49
2.4 Results ...... 51
2.5 Discussion ...... 56
2.6 Conclusion ...... 59
Chapter 3 An Automatic Method for PET Target Segmentation Using a Lookup Table Based on Volume and Concentration Ratio ...... 61
3.1 Abstract ...... 61
3.2 Introduction ...... 63
3.3 Methods and materials ...... 66
3.4 Results ...... 73
3.5 Discussion ...... 86
3.6 Conclusion ...... 94
Chapter 4 Conclusions and Suggestions of Future Work ...... 96
4.1 Conclusions ...... 96
4.2 Future work ...... 97
- 2 -
4.2.1 Validation of CT-PET image fusion ...... 98
4.2.2 Extensive application of the fiducial device ...... 99
4.2.3 Validation of the machine independence of the
new PET image segmentation method ...... 100
4.2.4 Application of the automatic PET segmentation
results for clinical outcome improvement ...... 101
4.3 Summary ...... 102
Bibliography ...... 104
- 3 -
List of Tables
Table 1.1 The physical properties of selected radionuclides used
in PET imaging system (17) ...... 25
Table 2.1 Fiducial registration error of fiducial based CT-PET
image fusion for each patient...... 52
Table 2.2 Average target registration error ± STD (mm) of
anatomical landmarks vs. fusion methods...... 53
Table 2.3 The results of one-way ANOVA F-test for the target
registration errors using manual, fiducial, and
automatic image fusion methods for each anatomical
landmark...... 55
Table 3.1 Threshold lookup table consists of target volume and
radioactivity concentration ratio. A desired threshold
for target segmentation is chosen from this table based
on the initial estimate of target volume and
concentration ratio...... 76
Table 3.2 Recovery coefficient table. Concentration ratio (C)
was corrected and recovered for partial volume effect
based on initial estimate of target volume and
measured source/background (S/B) ratio...... 79
- 4 -
Table 3.3 Volume calculation results using current method for
spheres were imaged in Philips Allegro PET scanner.
Each sphere was scanned with 4 different
concentration ratios ranging from 3:1 to 12:1 in
two-fold redundancy...... 81
Table 3.4 Volume calculation results using current method for
spheres were imaged in Philips Gemini TF PET/CT
scanner. Each sphere was scanned with 3 different
concentration ratios ranging from 3:1 to 12:1...... 82
Table 3.5 Volume estimation uncertainty (% error) comparison of
the current method with other published methods (88,
90)...... 83
Table 3.6 Target volume calculation for clinical patient images
compared with the CT-defined GTV...... 85
- 5 -
List of Figures
Figure 1.1 Positron emission and annihilation ...... 15
Figure 1.2 Types of false coincidence event...... 18
Figure 1.3 The physical limitation of the spatial limitation in PET
imaging system due to positron range and
non-collinearity effect...... 22
Figure 2.1 The design of the fiducial board...... 47
Figure 2.2 The set up of fiducial board on couch ...... 48
Figure 3.1 Workflow of the current PET target segmentation
method. First, a mean intensity method was used to
obtain the initial volume and S/B ratio (steps 1, 2, 3a,
and 3b). Next (steps 3c and 3d), the concentration
ratio was obtained by applying the recovery coefficient
table (Table 3.2). Based on the initial volume and the
recovered concentration ratio, a desired threshold was
then selected from the threshold lookup table (step 4).
Lastly, this threshold was used to perform a standard
level set method to delineate and estimate the final
target volume (steps 5 and 6)...... 70
Figure 3.2 Optimal thresholds yielding correct target volume for
- 6 -
each sphere target with different volume and
radioactivity concentration ratio in Allegro scanner (a)
and Gemini TF scanner (b)...... 74
Figure 3.3 Plots of S/B ratio versus measured sphere volume for
the phantom scanned in Allegro (a) and Gemini TF
scanner (b)...... 78
Figure 3.4 Comparison of CT volume and the segmented PET
volume for lesion No. 2. The left panel is the axial CT
slice with the manually defined GTV (green). The
right panel is the co-registered PET slice with the
automatically delineated contour as applied by the
current method (pink)...... 84
- 7 -
Acknowledgements
First of all, I would like to specially thank Dr. Barry W. Wessels, my research advisor, for his support, help and mentorship. I am extremely grateful for everything he taught me in work and in life. I could not finish this dissertation without his guidance and patience throughout my graduate study.
I would like to express my sincere thanks to other committee members,
Drs. Andrew M. Rollins, Xin Yu, and Syed F. Akber for their precious time, critical comments, and helpful suggestions. The same gratitude is also extended to the former committee members, Drs. Marty Pagel and Joseph
Syh.
I would like to thank all of my co-workers and colleagues in radiation oncology department, UHCMC. It is a pleasure to work with you all. The friendship, the training, the knowledge I received will benefit me for long time.
I would like to extend a special thank to Dr. Ravi Kulasekere for his warm-hearted advice on medical physics and generous help with my clinic duties all the time.
I deeply appreciate the support from my parents and my family.
Their love and understanding have been my most powerful resource. I wish
- 8 - my father could see this dissertation.
- 9 -
List of Abbreviations
CBCT: Cone Beam Computed Tomography
CI: Confidence Interval
CT: Computed Tomography
CTV: Clinical Target Volume
EBRT : External Beam Radiation Therapy
FBP: Filtered Back Projection
FDG: Fludeoxyglucose
FLE: Fiducial Localization Error
FOV: Field of View
FRE: Fiducial Registration Error
FWHH: Full Width at Half Maximum
GTV: Gross Tumor Volume
IGRT: Image Guided Radiation Therapy
IMRT: Intensity Modulated Radiation Therapy
LINAC: Linear Accelerator
LOR: Line of Response
MLC: Multi-leaf Collimator
MRI: Magnetic Resonance Imaging
MVCT: Megavoltage Computed Tomography
- 10 -
PTV: Planning Target Volume
RMS: Root Mean Square
RSS: Root Sum Square
TPS: Treatment Planning System
TRE: Target Registration Error
OAR: Organ at Risk
OSEM: Ordered Subset Expectation Maximization
PET: Positron Emission Tomography
PMT: Photomultiplier Tube
SPECT: Single Photon Emission Computed Tomography
- 11 -
CT-PET Image Fusion and PET Image Segmentation for Radiation Therapy
Abstract
by
YIRAN ZHENG
PET imaging system delivers abundant functional information which is complementary to the anatomical information provided by CT images. The purpose of this research is to improve the physician's ability to localize and delineate the extent of the tumor by incorporation of the PET images into radiation therapy treatment planning. A machine-based CT-PET fiducial fusion method was implemented for head and neck carcinoma radiation therapy. In this method, the field arrangements are aligned relative to the fixed treatment machine isocenter and patients are imaged in actual treatment positions. A fiducial registration error (FRE) of 1 mm was found for this fiducial fusion method. The target registration error (TRE) of seven anatomical landmarks was measured to evaluate the accuracy of this method.
The results were compared with a manual and a mutual information based automatic fusion method. Statistical analysis showed there was no
- 12 -
significant difference of TREs between the fiducial fusion method and the manual method which is considered to be most accurate in this research. In
addition, a new thresholding PET image segmentation method was proposed
using a lookup table which consists of the recovered activity concentration
ratios and the initial estimates of target volume. To validate the proposed
segmentation method, a Jaszczak phantom containing hollow spheres with
variable size and FDG concentration contrast ratio was scanned in different
PET scanners. The average uncertainty of the volume estimation by the
proposed method was 11.2% for spheres greater than 2.5 mL, which were
comparable or superior to those determined by contrast-oriented method and
iterative threshold method (ITM). This new segmentation method was also
applied to the PET images of ten patients with solitary lung metastases. The
average segmented PET volume was within 8.0% of the CT volumes. These
combined methodologies as outlined above are expected to decrease the
conformality index of the tumor dose (tumor volume/target volume) and spare
the normal tissue, which will result in an overall improvement in the effective
delivery of therapeutic radiation to patients. The suggested future work
includes further validation of the proposed methods at different PET scanners
and clinical application of these methods.
- 13 -
Chapter 1
An Introduction to PET and Radiation
Therapy Treatment Planning
1.1 Physics of positron emission tomography (PET)
Positron Emission Tomography (PET) is a nuclear medicine imaging
system used to create a three dimensional image of the emitted positron
within the human body. A positron is the anti-particle of electron, a
positively charged electron, emitted from unstable proton-rich radionuclides
which can occur naturally or be produced artificially using a cyclotron or
generator (1, 2). Positron emission is characterized as beta-plus decay. The
weak interaction converts a proton into a neutron and emits a positron and a
neutrino simultaneously (see Equation 1.1). Similar with beta decay, the
excessive decay energy is distributed between positron and neutrino.
Therefore, the emitted positron has an energy spectrum from zero to a
maximum value.
11+ 10p→ ne ++ν (1.1)
The emitted positron will lose energy via Coulomb interactions with the electrons or nuclei in the surrounding matter. A positron travels a short distance (~1mm) and will annihilate with an electron in the medium when the
- 14 -
positron reaches near thermal energy. This distance is called positron range, which depends on the initial kinetic energy of the emitted positron. The effect of positron range on the spatial resolution of PET imaging system will be discussed later in section 1.1.1. The positron-election annihilation will produce two identical photons, with same energy of 511 keV, traveling in opposite direction (see Figure 1.1). Around 40% of these high energy annihilation photons will penetrate through the human body and reach the detectors (3).
Figure 1.1 Positron emission and annihilation
(Adapted from http://gecommunity.gehealthcare.com/geCommunity/europe/
nmpet/pet_education/physics.html)
- 15 -
However, to preserve the residual kinetic momentum of the positron and electron before the annihilation, the two annihilation photons with same
energy are not strictly collinear. The FWHM (full width at half maximum) of
the angle deviated from 180° is about 0.5° (4, 5).
In the PET imaging system, the 511 keV annihilation photons will interact with the crystal scintillators surrounding patient in form of photoelectric effect and Compton scattering. These interactions will produce visible light (optical photons, ~1 eV) by scintillation. The amount of the
scintillation light is proportional to the number of the detected annihilation
photons. Then an adjacent photomultiplier tube (PMT) collects the light and generates electrical signals ready for processing. The single energy of the annihilation photons simplifies the design of the photon detection system.
Given the light speed and the size of the detector ring, the photons from the same annihilation event will reach the detector ring virtually at same time. This coincidence in time links the two detectors which interact with two annihilation photons independently. The line connecting these two linked detectors is known as the line of response (LOR). As long as two detectors capture the coincidence event, the corresponding annihilation must have originated somewhere on the LOR. LOR can be used to define the
position of a projection data, which is the equivalent function of collimation
but without a physical collimator. Therefore, it is called “electronic
- 16 -
collimator”. Electronic collimator has intrinsic advantage compared to the collimation in SPECT (Single Photon Emission Computed Tomography) which needs physical collimator to improve the spatial resolution and image quality
(6). A coincidence time window is set up to define the “coincidence” and this time window is about 10 nanoseconds (ns) for a conventional PET system.
A recent developed technique called time of flight (TOF) can achieve the time resolution better than 0.5 nanosecond (7). Better time resolution of detection can help to limit the location of the annihilation events to a line smaller than the diameter of the detector ring. The diameter of a detector ring is around 90 cm in clinical application, while photon travels 15 cm in 0.5 ns. The effective LOR with TOF technique is shorter than the conventional
LOR. Therefore, the contrast-versus-noise ratio and the image quality can be improved.
Not all coincidences detected by the photon detection system are true coincidence. The raw data produced by the photomultiplier tube (PMT) in the PET imaging system needs a number of corrections before and/or during the image reconstruction procedure. The typical corrections include attenuation correction, scatter correction, normalization correction, and random correction.
Attenuation correction is the most important step of data correction.
The interpretation of the PET image can be significantly changed without
- 17 -
attenuation correction. Depending on the location where the annihilation
occurs, one or both annihilation photons may be absorbed by body attenuation which results no LOR due to the lost event, see Figure 1.3 left panel.
Figure 1.2 Types of false coincidence event.
The probability of the photon attenuation, i.e. the amount of the
attenuated photons is related to the linear attenuation coefficient and the
body thickness on the true LOR. Basically, attenuation correction can be
performed by generating a three-dimensional linear attenuation coefficient
map in the whole scanned body. Therefore, the expected attenuation of any
pair of annihilation photons can be calculated as the linear integral of the
attenuation coefficient on the LOR. Conventionally, the attenuation
- 18 -
coefficient map is obtained by the transmission scan generated by an orbiting
68Ge/68Ga positron rod source outside of the scanned body. The principle is
exactly the same as a CT scan except the source is not an X-ray tube since the
PET detection system is optimized for photons with energy of 511 keV (the
energy of x-ray photons usually are below 150 keV). The recent development
of PET/CT combination scanner enables to use the CT image as the
attenuation coefficient map (8, 9). CT-based attenuation correction is
superior to positron-transmission-based attenuation correction because a CT
scan is much faster, and CT images have lower statistical noise comparing to
positron transmission images. However, some artifacts still exist, such as
motion artifact due to the organ motion between the scans and metal artifact
due to the implanted high density material (10).
For 511 keV photons, the photoelectric effect is negligible and the
Compton scattering is dominant in water. In the human body, the
attenuation is mostly caused by Compton scattering since the body’s density
and effective atomic number are both close to water. The scattered photon
may also reach one of the detector, see Figure 1.3 middle panel. A false LOR
will be formed since the scatter photons are not in the original direction of the
annihilation photon. Scattered coincidence event will decrease the PET
image contrast and modify the image interpolation. Scatter correction can be
performed by multiple energy window since these scattered photons usually
- 19 -
have a lower energy than 511 keV, or by model-based convolution/direct calculation (11, 12).
Random coincidence occurs when two photons from different annihilation events are detected within the predefined coincidence time window, see right panel in Figure 1.3 (13). Random correction can be
performed by the subtraction of a low-noise random coincidence estimate after
the data acquisition. Another common method of random correction is the
subtraction of a delayed coincidence signal during the data acquisition. Both
methods have trade-offs between noise, bias, and computation cost.
Another important correction accounts for detector response
normalization. Each individual detector has a different efficiency for
converting the incident photon to an electrical signal. Detector efficiency is
strongly affected by the detector geometry and the incident angle of photons
which is determined by the LOR position (14). The normalization factors for
normalization correction can be obtained by scanning a uniform planar or
cylinder source (15). In addition, the PET acquisition data also needs
correction for detector dead time, multiple-coincident, source decay.
After the raw data is corrected, a projection based image reconstruction algorithm such as Filtered Back Projection (FBP) or Ordered
Subset Expectation Maximization (OSEM) is used to reconstruct the PET image, which is similar with CT or SPECT image reconstruction. Interested
- 20 -
readers are referred to Chapter 4, “Image Reconstruction Algorithms in PET”
in book of “Positron Emission Tomography”, Springer London, 2005, for technical details.
1.1.1 The spatial resolution of PET imaging system
One of the major concerns about PET imaging is its relatively low
spatial resolution. The spatial resolution of an imaging system is defined as
the minimal distance between two lines or point sources which can be
distinguished or resolved as separate details in image. An alternative
definition of spatial resolution is the FWHM of the imaging system’s point
spread function (PSF). Although pixel size of the reconstructed image can be
interpolated to any value, the spatial resolution is limited by the imaging
system properties. Recent commercially available PET systems used
clinically have spatial resolutions on the order of 4–7 mm (16), which is much
worse than those of CT and MRI.
Unlike other imaging systems, the spatial resolution of PET is limited
by two inevitable physical effects: the positron range before annihilation and
the non-collinearity effect caused by the residual kinetic momentum of the
electron-positron pair at annihilation (4). These two effects cause the
deviation between the true LOR and the observed LOR (see Figure 1.3).
- 21 -
Figure 1.3 The physical limitation of the spatial limitation in PET imaging system due to positron range and non-collinearity effect.
Adapted from http://kakuigaku.cyric.tohoku.ac.jp/PETAnalysis/PET_ABC_v2-1.pdf
The positron range depends on the initial kinetic energy of the positron. The mean positron range in water for 18F, the most widely used positron emitter in the clinic, is about 0.6 mm (17).
The deviation from the collinearity has a FWHM of 0.5° (4). The related FWHM of resolution blurring can be approximately estimated by
Equation 1.2, where D is the detector ring diameter. For example, given a
- 22 -
detector ding diameter, D, equals to 90.34 cm in Philips Gemini TF PET/CT scanner (18), the FWHM of the spatial resolution degradation due to non-collinearity is 2.0 mm.
0.5 π D FWHM − ≈×( ) = 0.0022D (1.2) non col 2 180 2
Another major factor influencing PET spatial resolution is the size of detector because a detector element (a single scintillator) cannot give the
information pertaining to the location of the entrance of a given incident
annihilation photon. Since theoretically the detector sensitivity is maximum
for the LOR at the center and falls off to zero at the edge, detector size effect is
modeled as a triangular response function (19), assuming the point source is
near the center of the FOV (field of view). The corresponding FWHM of the
blurred resolution equals to half of the detector element width. The
resolution degradation is worse at the edge of the scanner FOV. The typical
detector size of a clinical PET system is around 6 mm, and the related FWHM
of resolution blur is about 3 mm.
Besides the factors mentioned above, the spatial resolution of PET
system is also affected by detector layout geometry, depth of interaction
within the detector element, and the image reconstruction algorithm. The
- 23 -
reconstruction algorithms usually trade off spatial resolution for noise reduction and a factor of 1.2 to 1.5 on resolution degradation is induced depending on the algorithm (4). The overall spatial resolution is a convolution of all these factors, and can be estimated as the root-sum-square
(RSS) of each resolution blur FWHM resulted from different factors.
1.1.2 18F-FDG - the radiotracer used in PET
PET can provide a functional or metabolic assessment of normal tissues or neoplasms by imaging the concentration and distribution of specific
chemical compound labeled by positron-emitting radionuclides in the body
(20-22). The common labeling radionuclides are Carbon-11, Fluorine-18,
Oxygen-15, and Nitrogen-13.
The physical properties of common radionuclides used in the current
PET imaging systems are listed in Table 1.1. According to Table 1.1, 18F has the shortest mean positron range in water which results in better spatial resolution of PET image. The appropriate half-life of 18F is also preferred in
clinical application. The use of a radionuclide with a longer half-life results in higher total radiation dose received by patient because it will take longer time for radionuclide to decay. A radionuclide with a shorter half-life has the advantage to increase the signal-noise-ratio since the administration dose can
be higher, assuming the total radiation dose is same. However, the higher
- 24 -
administration dose (the dose injected into patient’ body before the scanning) causes possible technical difficulties in the production, transportation, and injection of the radiotracer.
Table 1.1 The physical properties of selected radionuclides used in PET imaging system (17)
Maximum Mean positron half-life Isotope positron energy range in water (min) (MeV) (FWHM in mm)
11C 20.3 0.96 1.04
13N 9.97 1.19 1.31
15O 2.03 1.70 2.00
18F 109.8 0.64 0.61
68Ga 67.8 1.89 2.21
82Rb 1.26 3.15 4.21
18F labeled FDG (18F-fluoro-deoxyglucose) is the most widely used radiotracer in clinical PET application and the only oncologic PET tracer
- 25 -
approved by FDA for routine clinical use. More than 90% of PET
examinations is based on 18F-FDG which can be produced in on-site cyclotrons
(23). FDG is an analogue of glucose, which enters cells via the same mechanism as that of glucose. Tumor cells can be differentiated from normal cells at a molecular and biochemical level which increases their glucose uptake and therefore affects their metabolic rate. Therefore, the tumor cells accumulate 18F-FDG more avidly than normal cells due to the higher
metabolism level, which has given rise to the potential of PET imaging in
oncology (24). Recent literature shows FDG-PET improves the staging,
treatment response, restaging after therapy and prognosis for lymphoma (25),
lung cancer (26, 27), head and neck tumors, (28), thyroid carcinoma (29),
breast cancer (30), ovary cancer (31), sarcoma (32), and other malignances.
In addition, the investigation shows that the change of 18F-FDG uptake measured by PET may provide useful information of subclinical response of tumors to cancer therapy (33).
- 26 -
1.2 Radiation therapy treatment planning
1.2.1 Radiation therapy and External Beam Radiation Therapy
Radiation therapy or radiotherapy is the application of ionizing
radiation to control malignant or benign tumors while sparing the normal
tissues. Radiation therapy procedures can be classified into two general
categories, brachytherapy and external beam radiation therapy (EBRT),
which is also referred to as teletherapy.
In brachytherapy, the radiation dose is applied in close proximity to
the intended target volume. The radioactive material is placed into the
target volume (intracavitary or interstitial brachytherapy) or directly on the
target (surface mould or intraoperative radiotherapy) (34, 35). These sources
have a short range of radiation penetration which aids in minimizing the
unwanted dose to the adjacent normal tissue. Some examples of
radionuclides used in brachytherapy are 125I are 103Pd, which are used in low
dose rate brachytherapy, and 192Ir, which is used in high dose rate brachytherapy. The physical advantage of brachytherapy treatments is the
localized delivery of the radiation dose to the target volume of interest (35).
However, invasive operation must be performed by physician in
brachytherapy. Furthermore, brachytherapy can only be used for localized tumor with relatively small volume.
- 27 -
In external beam radiation therapy, the radiation source is at a
certain distance from the treatment target. Typically, a megavoltage linear
accelerator (LINAC) or a cobalt-60 unit is used to generate the external radiation beam. Most external beam radiation therapy treatments are carried out with photon beams and electron beams. It should be noted that
heavy particles such as protons, carbon ions, and neutron are also used in
some therapy applications. The cobalt-60 unit only delivers a photon beam
with average energy of 1.25 MeV. A linear accelerator (LINAC) usually
delivers multiple energy photon and electron beams which provides flexibility
to treat target volumes with complex shapes located at a variety of depths.
Generally, there are two basic methods for patient setup in
conventional external beam radiation therapy: SSD (source-to-surface
distance) setup and SAD (source-to-axis distance) setup. In a SSD setup, the
distance between the radiation source and patient’s skin is the reference
distance in treatment plan and recorded in chart. In a SAD setup, the target
volume is placed at a virtual center point (isocenter) of the treatment unit.
The treatment machine isocenter is defined as the point of intersection of the
gantry rotating axis and the beam axis. The SAD setup is appropriate for
isocentric radiation therapy, which delivers radiation dose from multiple
angles to the isocenter. For a given patient, the location of the isocenter is
essential during all the isocentric radiation therapy procedures including
- 28 -
simulation, treatment planning, patient positioning, and radiation delivery.
The treatment machine isocenter must be associated with the plan isocenter
defined in the treatment plan.
In radiation therapy, the objective is to deliver a high radiation dose to
sterilize all tumor cells and while minimizing the radiation damage to the surrounding normal tissues. Radiation beam scattering originating in the treatment machine and within the patient is one of the reasons that it is difficult to achieve such objective when the normal tissues, especially the radiation sensitive organs or organs at risk (OAR), are close to the target volume. To improve the therapeutic value of radiation therapy, high-dose conformal radiation therapy is preferred. The aim of the conformal radiation therapy is to deliver a high dose to the target volume while minimizing the dose to OARs and the other surrounding normal tissues. The conformality index can be used to describe how well the dose distribution covers the target volume, which is defined as the ratio of target volume to tumor volume.
Typically, three-dimensional conformal radiation therapy (3D-CRT) is achieved by delivering a few static radiation beams from different angles.
The shape of the beam can be determined by the projection of the target volume at beam’s angle. The intensities across the field of these radiation beams are usually uniform and tailored by simple beam modifiers, such as wedges or compensators. 3D-CRT technique has evolved into intensity
- 29 -
modulated radiation therapy (IMRT) to accomplish the task for more complex
target and OAR geometries. Using advanced computer-controlled
megavoltage linear accelerators with multi-leaf collimators (MLC), intensity modulated radiation therapy (IMRT) uses optimized non-uniform beam
intensities from multiple angles to obtain a uniform high dose in tumor
volume and minimize the dose to the surrounding normal tissues (36). IMRT
treatment planning is usually performed by sophisticated optimization
algorithm based on the preset plan objective or desired dose distribution.
Furthermore, the recent developed Tomotherapy unit can deliver a continuous
modulated helical radiation dose to the target volume to achieve the purpose
of intensity modulated radiation therapy (37).
Patient immobilization and positioning is important and requires
special attention in conformal radiation therapy (38, 39). In highly
conformal radiation therapy, since the target does is escalated and the dose
decreases sharply at the edge of the target to spare the surrounding normal
tissues, the target volume or patient must be placed at the expected location
with high spatial accuracy. Customized and patient-specific immobilization
devices can immobilize the patient during treatment and provide a reliable
method to reproduce the patient’s position from simulation to treatment and
from one treatment to another. The same immobilization device and setup
should be applied in both simulation and treatment. Thermal plastic masks
- 30 -
or whole body moulds (thermal plastic mould, Alpha Cradle or Vac-Lok) are
commonly used as immobilization devices (40). If patient lost weight or the
tumor shrinks in the middle of the treatment course, oncologist may consider
a new simulation, new immobilization device, and new treatment plan to
adapt the change.
During the irradiation, patients are usually positioned on the treatment couch by laser alignment. In addition, target volume position and patient position on treatment couch should be verified before the treatment.
Conventionally, patient position verification is performed by portal imaging, which uses the radiation source to generate a radiographic film of patient’s
position on the treatment couch. Orthogonal portal imaging is preferred
since the film is in 2D format. Recently developed image-guided radiation therapy (IGRT) techniques use on-board imaging system to verify the position of target volume (41, 42). One example of megavoltage computed tomography (MVCT) can be found on the Tomotherapy unit which can generate a 3D image volume using the helical radiation beam when patient is positioned on the treatment couch (43). The MVCT image can be registered with the stored planning CT. The difference between plan and actual position can be compensated by moving the couch. Cone-beam computed tomography (CBCT) using the cone-shaped radiation beam can also accomplish the same objective of target position verification (44).
- 31 -
Image-guided radiation therapy (IGRT) is expected to deliver radiation dose
the target volume more accurately than the conventional radiation therapy
and to reduce the uncertainty during the radiation therapy (45).
1.2.2 Treatment planning for radiation therapy
Before delivering radiation dose to patient, a treatment plan needs to be generated at a treatment planning system (TPS). A treatment plan specifies the patient’s setup, beam angle, beam shape, use of beam modifier, treatment time or monitor unit (MU) for each field/segment, etc, needed to satisfy the radiation oncologist’s dose prescription to target and dose restrictions on organs at risk (46). A treatment plan also generates a
three-dimensional dose profile or dose-volume histogram (DVH) for plan
evaluation. Treatment planning includes many steps such as patient data
acquisition and simulation, volume definition, treatment time or monitor unit
(MU) calculation, optimization, plan evaluation, quality assurance.
The purpose of the simulation is to determine the patient’s treatment
position, identify the target volume and organs at risk, verify treatment filed
geometry, and also collect patient data. Simulation is usually performed
with an X-ray based conventional simulator or a CT simulator which is widely
used in current clinical application. Technically, a CT simulator is the same
as a diagnostic CT scanner with following two exceptions: the
- 32 -
laser/mechanical localization system and the flat couch top which reproduces
the couch top of the treatment machine. Usually, the couch top of the
diagnostic CT scanner is curved for patient’s comfort (47).
Based on the simulation data, volume definition is required for
meaningful treatment planning and accurate dose reporting. Organs at risk
(OAR) and tumor volumes such as gross tumor volume (GTV), clinical target
volume (CTV), and planning target volume (PTV) need to be delineated before beam optimization and dose calculation. According to ICRU (International
Commission on Radiation Units and Measurements) Report No. 50, “The
Gross Tumour Volume (GTV) is the gross palpable or visible/demonstrable
extent and location of malignant growth”. Gross tumor volume (GTV)
delineation is essential to the entire radiation therapy process and is usually based on the information obtained from a combination of imaging modalities
such as computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET).
The clinical target volume (CTV) is GTV plus the tissue volume containing sub-clinical microscopic malignant disease and/or the tissue
volume considered to be at risk and requiring treatment. The degree of
extension beyond GTV to CTV is primarily decided by radiation oncologist
according to clinical examination, pathology, and experience. Planning
target volume (PTV) is CTV plus additional margin to eliminate the effect of
- 33 -
possible geometrical uncertainty on delivering the prescribed dose to CTV.
PTV is associated to the treatment machine coordinate system and can be described as the CTV plus certain distance (e.g. PTV = CTV + 5 mm). The appropriate margin is based on systematic and random uncertainties in patient positioning system, immobilization device, patient position verification method, internal organ motion, and respiratory motion. The margin can be reduced using image-guided radiation therapy (IGRT) techniques to spare the surrounding normal tissue more efficiently. The organ at risk (OAR) is an organ which may receive a considerable dose from the treatment plan compared with its radiation tolerance, etc. Most OARs
are adjacent to the treatment target or radiation sensitive organs with low
tolerance dose.
Besides the delineation of the tumor volumes and organs at risk,
oncologists also specify the prescription dose to target, dose limitation to the
OARs, fraction size. These physician’s prescription and orders are set as the
objective in treatment planning system (TPS). The optimization algorithms
may adjust the beams arrangement and relative dose weight automatically to
fulfill the objective by an iterative procedure. After the objective is achieved,
treatment planning system (TPS) generates dose-volume histogram (DVH) or
3D dose distribution profile, and physicist/physician evaluates the plan.
Additionally, a secondary check by direct measurement or software other than
- 34 -
the TPS is required to be performed for quality assurance purposes.
1.3 The role of PET in radiation therapy treatment planning and the
significance of this research
FDG-PET imaging system provides functional information of metabolism which is complementary to the anatomical information provided by CT images. Investigation shows that FDG-PET has an important role in diagnostic and assessment of several neoplasms (48). Besides its use as a diagnostic tool in oncology, PET is increasingly being used as a planning tool for radiation therapy, specifically to define the target volume (49). The interobserver and interobserver variation of GTV definition can be reduced by the image registration of diagnostic PET images and planning CT images, comparing with using CT or PET images alone (50, 51).
Usually, the planning CT taken from a CT simulator is used as a reference frame in the current clinical applications. Any other images involved in the treatment planning need to be “fused” with the planning CT to map the treatment machine coordinate system. Strictly, the terminology of
“image fusion” used in clinic, and in this dissertation, includes two steps, image registration and image fusion. Image registration is geometrical alignment by transforming different image data sets into one coordinate
- 35 -
system. The images can be from the same or different imaging modalities.
Image fusion is a procedure of superposing two registered images using simple
mathematical operation, such as addition, subtraction, and maximum. The quality of image fusion depends on accurate image registration. For images from same modality, checkerboard fusion is the common technique used in clinic. Checkerboard fusion splits the image into different regions which alternate between two different image volumes. For images from different modalities, image fusion is usually performed by RGB addition method. In the fused image by RGB addition, the displayed color at any pixel is a scaled summation of the pixel colors from the two different image volumes which may have different color maps, such as black-white, thermal, blue and so forth.
In addition, other mathematically operations for fused image can be applied for different purpose. For example, subtraction fusion subtracts the pixel value of the secondary image from the primary image to enhance the difference between the two fused images.
In radiation therapy treatment planning, CT-PET image fusion is usually performed by manual alignment of anatomical structures. Fiducial fusion based on the fiducials attached on patient’s surface or implanted in patient’s body is another reliable option. Mutual information-based automatic image fusion can be used as assistant tool, which requires a reasonable starting point and a manual adjustment/confirmation to be
- 36 - followed. However, low spatial resolution, partial volume effect and lack of anatomical details limit the operator or algorithm’s ability to fuse PET and CT image quickly and accurately. Furthermore, the coordinate system in diagnostic PET image is not related to treatment machine reference frame currently. Although PET/CT combination scanner can provide a “hardware approach” to image fusion (52), operational difficulty still remains for CT-PET image fusion used in radiation therapy treatment planning since, most of the time, the planning CT is performed separately from the PET study. This is especially true when a patient has a diagnostic PET coming in from an outside institution. To ensure that the diagnostic PET image is fused accurately with planning CT and related to treatment machine, a machine-based fiducial fusion method was implemented for head and neck carcinoma in radiation therapy treatment planning system in this research. With the immobilization equipment and fiducial device, patients are imaged in actual treatment positions and the fusion and field arrangements are aligned precisely relative to the treatment machine isocenter. The clinically relevant result provided by this method is expected to assist in patient positioning and improve the quality of radiation therapy.
Given that the CT-PET image fusion is completed accurately, the functional information carried by the PET image can improve the gross tumor volume (GTV) definition along with CT. The process of target delineation is
- 37 -
also called image segmentation. Although there are many sophisticated
automatic or semi-automatic image segmentation techniques available for CT
or MR images, thresholding method is still widely used in PET image
segmentation due to the low spatial resolution and high partial volume effect.
However, the desired optimal threshold depends on target size and
target-to-background ratio significantly (53). In order to improve on current
limitations of thresholding methods and to provide a rapid and convenient
method for PET target volume delineation, a combined multi-step
thresholding and level set segmentation method was developed in this research. The experimental results from phantom and patient data studies demonstrate this method to be robust and machine independent with adequate accuracy for wide range of target volumes and source-to-background
(S/B) ratios.
1.4 Overview of dissertation organization
In Chapter 2, a machine-based CT-PET fiducial fusion method is introduced. This method is based on a fiducial board embedded with a “Z” shape tube visible in both CT and PET images. Ten patients with head and neck cancer were imaged in CT and PET scanner with this fiducial board.
For purposes of comparison, manual, machine-based fiducial and mutual
- 38 -
information-based automatic fusions were completed. Fiducial registration
errors for fiducial markers and target registration errors were measured for
anatomical landmarks to assess the accuracy of clinically applied PET and CT
image fusions for planning and delivering radiation therapy. Statistical
analysis showed that the differences in target registration errors for manual,
fiducial, and automatic fusion were not significant.
In Chapter 3, a method for rapid and accurate PET image
segmentation and volumetrics based on phantom measurements and
independent of scanner calibration is introduced. It has been confirmed that the optimal segmentation thresholding depends on target volume and radioactivity concentration ratio. A three-step method based on the PET image intensity information alone was used to delineate volumes of interest including target volume initial estimation and target-to-background concentration ratio recovery. The segmentation results were compared with those determined by contrast-oriented method and iterative threshold method
(ITM). In addition, the new volume segmentation method was applied clinically to ten patients undergoing PET/CT volume analysis for radiation therapy treatment planning of solitary lung metastases.
Finally, in Chapter 4, major conclusions and significance of this research are summarized. It is also discussed about the possible future
validations (4 items) including the developments and clinical applications of
- 39 -
the machine-base fiducial fusion method during treatment and using the new thresholding PET image segmentation on different PET scanners (Section
4.2).
- 40 -
Chapter 2 A Machine Based CT-PET Image Fusion System
2.1 Abstract
Although PET/CT combination scanner can provide a “hardware approach” to image fusion, operational difficulty still remains for PET-CT image fusion used in radiation therapy treatment planning since the planning
CT much of time is performed separately from the PET study. This is especially true where a patient has diagnostic PET coming in from an outside institution. This study is to assess the accuracy of clinically applied PET and
CT image fusions for planning and delivering radiation therapy using manual, fiducial-based, and automatic registration techniques. A machine-based fiducial device was developed and used to co-register patient and machine isocenter. This fiducial device is a lucite board (61 cm by 38 cm by 5 cm) with
0.6 cm diameter tube in the shape of a “Z” embedded. The tube was filled with contrast (10% Optiray) for CT scans and 18FDG (100μCi/22 ml) for PET imaging. This fiducial board was mounted directly under the patient on CT or PET scanner couch during the imaging procedures. Ten patients with head and neck cancer were locked on the fiducial board with a thermoplastic mask and a treatment board, and received simulation CT scan and diagnostic
PET scan on same day. The fiducial board was invariant for both CT and
- 41 -
PET scans and so considered the “gold standard” for accurate treatment
delivery in relation to isocenter. Fiducial fusion was performed for each
patient by aligning the fiducial line markers which were visible on both CT
and PET images. For comparison, manual and mutual-information-based automatic fusions were also completed. After image fusion, fiducial registration errors for fiducial markers and target registration errors for seven anatomical landmarks (anterior tip of chin, anterior tip of nose, superior tip of scapula, anterior edge of frontal lobe, base of skull, center of the
C4 vertebrae and anterior center of mandible) encompassed the treatment site were measured. The average fiducial registration error was 1.07 mm. The small fiducial registration error demonstrated that the fiducial board can be used as the reference for patient treatment position relative to the machine isocenter. One-way ANOVA statistical analysis showed that the differences of target registration errors for manual, fiducial, and automatic fusion were not significant at 95% confidence interval except anterior edge of frontal lobe
(p = 0.007) and base of skull (p = 0.043). The average target registration errors for the selected anatomical landmarks in all methods exceeded expected alignment uncertainties nominally associated with PET and CT image fusion techniques, although patient immobilization equipments were used during imaging procedures.
- 42 -
2.2 Introduction
Positron emission tomography (PET) is a noninvasive radiotracer imaging technique. Radiotracers are molecules labeled with a positron emitting isotope such as 18F, 11C, 13N or 15O, which allow the measurement of a
biological process in an organ or tissue and can provide functional/metabolic
information (54). [18F]2-fluoro-D-2-deoxyglucose (FDG), a glucose analogue, is the most commonly used radiotracer today and is readily available with a half life that is acceptable for clinical applications. Today, FDG-PET is used in a variety of clinical applications including cancer staging and patient follow up (48). FDG-PET improves the staging, treatment response assessment, restaging after therapy and prognosis for lymphoma (25), lung cancer (26, 27), head and neck tumors (28), thyroid carcinoma (29), breast cancer (30), ovary cancer (31), sarcoma (32), and other malignances. Besides its use as a diagnostic and staging tool in oncology, PET is increasingly used for target volume definition in radiation therapy planning (49). However, the ability of
PET to accurately delineate tumor from uninvolved tissue is limited by low spatial resolution, physiologic uptake by normal processes, partial volume effects, and lack of anatomical details. Further study is needed to define how best to incorporate PET imaging into radiation therapy treatment planning with the ultimate goal of improving patient outcomes.
- 43 -
While PET imaging lacks anatomic information which is essential for effective radiation treatment planning, PET-CT image fusion combining
anatomical and functional information has the potential to improve patient
management (55). PET/CT fusion images are more accurate than PET or CT alone for the staging and localization of malignances such as head and neck cancer (56) and non-small cell lung cancer (NSCLC) (57). Currently,
mutual-information-based image registration is most commonly used
technique for automated PET-CT image fusion (58). However, the automated
registration is often unsatisfactory and manual registration or adjustment of
the automated fusion is usually required. Furthermore, PET-CT data is
often captured with the patient in a diagnostic position which does not
correlate to the patient in a treatment position. More recently combination
PET/CT scanners have been developed which integrate the two imaging
modalities into one device and provide a hardware approach to image fusion
which is considered as true image fusion (52). However, combination
PET/CT scanners are limited by patient movement between the CT and PET
data acquisitions, misalignment, and organ motion. Also, for treatment
planning, the PET image needs to be fused with the planning CT image taken
in a dedicated simulation CT scanner.
Machine-based fiducial devices can improve image fusion and have
been employed in SPECT/CT, SPECT/MR (59), and for patient positioning for
- 44 -
GammaKnife radiosurgery. In order to exploit the superior fusion that can be achieved with machine-based fiducial devices, we designed a machine-based fiducial fusion method for radiation treatment planning for head and neck cancer patients. With the immobilization equipments and fiducial device, patients are imaged in actual treatment positions and the fusion and field arrangements are aligned precisely relative to a fixed isocenter. A machine-based fiducial device is expected to improve patient positioning and improve the quality of radiation therapy.
It is necessary to validate the clinical usefulness of this machine-based fiducial fusion method and estimate the registration error.
Besides PET/CT option, there is no independent gold standard to validate
PET-CT patient image registration unless certain fiducial markers are placed on patient surface or implanted in patient’s body which is not always available in clinical application (60). The investigator only can estimate the fusion images visually to determine the registration accuracy for particular patient.
To evaluate the accuracy of this machine-based fiducial fusion method, the target registration errors of many anatomical structure landmarks encompassing the treatment area will be measured by visual assessment.
The results will be compared with a manual registration technique and standard mutual-information-based automatic registration technique when applied to PET-CT image fusion for planning and delivering radiation therapy.
- 45 -
2.3 Methods and materials
2.3.1 Patients
A series of ten patients with head and neck cancer were studied. A machine-based fiducial device was used for patient’s PET-CT image fusion.
However, this device was not employed for patient’s radiation treatment in the current study. During treatment planning, the fiducial device and CT couch were both removed and replaced by the treatment machine couch in planning
CT image.
2.3.2 Image Acquisition
Central to this study is a fiducial system that can be registered to both the patient and machine isocenter and remained clearly visible on both CT and PET imaging. A lucite fiducial board was constructed that measured 61
cm x 38 cm x 5 cm which was locked to indexed tables (Figure 2.1).
- 46 -
Figure 2.1 Fiducial board design with Z-shaped embedded tube.
In the plane of the patient, a 0.6 cm diameter tube in the shape of a
“Z” was designed into the board which was later filled with either contrast
(Optiray® 240, 10% solution) for CT scans or 18FDG (100 μCi / 22 ml) for PET scans. Both the fiducial board plus any other auxiliary devices (e.g. mask,
IMRT board) were applied in patient CT and PET scanning. Patients underwent CT simulation scan (Philips AcQSim) and PET imaging (Siemens
ECAT) lying on the fiducial board on the same day. Customized pillow and
Aquaplast RT® thermal plastic mask were used as immobilization device
during both CT and PET scanning. Immobilization devices were fixed to the
- 47 -
fiducial board through the Med-tech board and locked into the indexed CT and
linac couches. Figure 2.2 demonstrates the fiducial system setup. Since the
PET couch is not indexed, customized lasers were used to guide the board position. The fiducial board was positioned underneath the treatment site and invariant between the CT and PET scans. The fiducial device did not affect CT or PET imaging procedures and the standard clinical imaging protocols were followed. The pixel size of the CT image was 1 mm by 1mm with a slice thickness of 3 mm. The image resolution of the reconstructed
PET image was 4 mm by 4 mm with a slice thickness of 4 mm.
Figure 2.2 Fiducial board setup.
- 48 -
2.3.3 Image Fusion
Following imaging, manual PET-CT fusion and a standard mutual-information-based automated PET-CT fusion were performed for each patient using MIMVista® software (rigid body registration, 6 degrees of freedom). In addition, we developed an algorithm for automated fiducial fusion using a Hough transform (61) which uses a feature extraction technique to identify lines in the images. This algorithm was used to perform a machine-based PET-CT fiducial fusion for each patient. The visible line in CT and PET images will be the “gold standard” in fiducial fusion method for accurate treatment delivery since it has a direct mechanical relationship to treatment machine isocenter. These fusions ensured the images used for planning were located accurately relative to each other for treatment planning. For subsequent treatment, the fusion and field arrangements must be aligned precisely relative to a fixed machine isocenter.
2.3.4 Data Analysis
Validation and evaluation of the machine-based fiducial fusion method
was performed by calculating the fiducial registration error (FRE) and the
target registration error (TRE) of anatomical structures which were measured
by visual inspection. Fiducial registration error (FRE) was defined as the
- 49 - distance between corresponding fiducial markers after PET-CT image fiducial fusion. Because the fiducial used in this study was not a standard point source fiducial, the FRE measurement was done by averaging the FRE of the visible fiducial markers in 5 axial slices around the patient’s treatment site.
Target registration error (TRE) was defined as the distance between corresponding points of anatomical landmark other than the fiducial markers on the fused CT and PET images. Seven anatomical structures surrounding treatment site were carefully chosen including anterior tip of chin, anterior tip of nose, superior tip of scapula, anterior edge of frontal lobe, base of skull, center of the C4 vertebrae, and anterior center of mandible. Three different methods of image registration were compared when applied to PET and CT image fusions used for planning and delivering radiation therapy. These three methods use a manual, a machine-based fiducial and mutual-information-based automatic registration technique. Statistic analysis was performed on the results of TRE measurements using three image fusion methods for comparison of the accuracy. One-way ANOVA
F-test was used to test the null hypothesis that there was no difference in registration accuracy among these three fusion methods. The F-value and the corresponding P-value were calculated for each anatomical landmark. If the one-way ANOVA F-test showed the registration accuracy resulted from the three fusion methods was significantly different for certain anatomical
- 50 -
landmark, two sample Student’s t-test was followed to exam the registration
accuracy difference resulted from each two of the three fusion methods.
2.4 Results
Table 2.1 shows the fiducial registration error for each patient’s
CT-PET image fusion using fiducial registration method. For ten cases, the average fiducial registration error was 1.07 mm and standard deviation was
0.68 mm. Fiducial/fiducial alignment error of around 1 mm characterizes the absolute accuracy achieved for small objects regular in shape, high contrast, and spatially stationary over time.
- 51 -
Table 2.1 Fiducial registration error of fiducial based
CT-PET image fusion for each patient.
Fiducial registration Patient No. error (mm)
1 0.34
2 0.51
3 1.48
4 0.73
5 0.81
6 1.11
7 0.44
8 2.58
9 1.53
10 1.20
Average 1.07
STD 0.68
- 52 -
Table 2.2 summarizes the experimental results of the target registration error measurements over 10 patients’ CT-PET image fusion using manual, machine-based fiducial and mutual information based automatic registration techniques. Seven anatomical structures around target area were chosen to evaluate the accuracy of three image fusion methods. These errors associated with the anatomical landmarks represent patient’s shift and internal organ’s movement in location relative to the isocenter between or during CT and PET scans.
Table 2.2 Average target registration error ± STD (mm)
of anatomical landmarks vs. fusion methods
Average target registration error ± STD (mm)
Anatomical landmark Manual Fiducial Automatic
method method method
Anterior tip of chin 5.72 ± 3.58 7.14 ± 2.79 7.28 ± 3.17
Anterior tip of nose 7.57 ± 3.09 8.79 ± 3.28 7.92 ± 4.18
Superior tip of scapula 10.82 ± 9.53 11.85 ± 8.63 6.44 ± 4.17
Anterior edge of frontal lobe 4.34 ± 0.81 5.20 ± 1.98 7.52 ± 2.70
Base of skull 4.65 ± 1.98 8.02 ± 3.19 7.42 ± 3.66
Center of the C4 vertebrae 5.35 ± 2.41 7.98 ± 2.98 6.24 ± 4.62
Anterior center of mandible 7.80 ± 3.40 10.09 ± 2.50 8.68 ± 3.09
- 53 -
One-way ANOVA statistical analysis (Table 2.3) showed that the differences of target registration errors for manual, fiducial, and automatic fusion were not significant at 95% confidence interval (CI) except anterior edge of frontal lobe (p = 0.007) and base of skull (p = 0.043). Two sample
Student’s t-test for anterior edge of frontal lobe showed there was no significant difference between the TREs using manual and fiducial fusion method at 95% CI, while the TRE using automatic fusion method was greater than manual fusion method (p = 0.004) and fiducial fusion method (p = 0.028).
For base of skull, there was no significant difference between the TREs using fiducial and automatic fusion method at 95% CI. However, the TRE using manual fusion method was less than fiducial (p = 0.006) and automatic fusion method (p = 0.028).
- 54 -
Table 2.3 The results of one-way ANOVA F-test for the target registration errors using manual, fiducial, and automatic image fusion methods for each anatomical landmark.
Anatomical landmark F-value P-value
Anterior tip of chin 0.73 0.490
Anterior tip of nose 0.32 0.732
Superior tip of scapula 1.36 0.274
Anterior edge of frontal lobe 6.15 0.007
Base of skull 3.54 0.043
Center of the C4 vertebrae 1.50 0.242
Anterior center of mandible 1.46 0.250
The registration error associated with the inherent uncertainty of the
CT-PET image fusion due to limited image resolution can be estimated as the root-mean-square (RMS) of the CT pixel size (1 mm) and the PET pixel size (4 mm). In this study, this uncertainty was found to be 2.92 mm. The average target registration errors for all anatomical landmarks and all of three image fusion methods exceeded this uncertainty (p < 0.02).
In addition to providing similar results to manual image fusion, fiducial fusion method significantly reduced the image processing time up to 5
- 55 -
minutes, comparing to 10 minutes on average for manual fusion in our
institution. The fiducial markers in the fused image incorporated the patient
treatment position and improved the efficiency in radiation therapy treatment
planning.
2.5 Discussion
In this study, a new CT-PET image fusion method using a machine-based fiducial device was developed for head and neck carcinoma radiation therapy. This special fiducial device is attached on the top of CT and PET couches during scanning. Because this device does not adhere to patient’s body or skin, and can be installed prior to the scanning, using this device will not affect the regular medical imaging procedure in clinic. In the current method, patient was immobilized by a thermal plastic mask and a treatment board which are standard devices in head and neck radiation therapy. The patient’s movement between different imaging scanning and treatment is minimized, although this movement cannot be eliminated.
Based on the alignment of the fiducial marker in both CT and PET image, this fiducial fusion method does not require identifying the anatomic landmark for image registration. Furthermore, an algorithm using Hough transform (61) can automatically detect the linear feature of the visible fiducial markers on both CT and PET image and enables to perform automatic fiducial fusion.
- 56 -
Using this fiducial device and the immobilization equipments, patients are
imaged in actual treatment positions and the PET image can be fused
precisely relative to the treatment machine isocenter.
Currently, PET/CT combination scanner is widely used in diagnosis,
staging and restaging of cancer patients for various tumor sites, which
improves the diagnostic interpretation of both PET and CT images and allow
CT-PET “hardware fusion” (62). However, the CT-PET image fusion remains
an issue for radiation therapy treatment planning since the planning CT is
usually performed separately in a CT simulator with a flat tabletop, laser localization and customized patient immobilization device (63). Image fusion
of planning CT scan and PET scan from PET/CT scanner cannot access the
benefit of PET/CT combination and the complexities are not resolved, which
limits the utility of PET/CT scanner for radiotherapy treatment planning.
Especially, when patient receives diagnostic PET scan at outside institution
and cannot take second PET scan due to financial consideration, the CT image
from PET/CT scanner is not applicable for radiation therapy treatment
planning. In order to establish a reference frame matching with treatment
machine, it is suggested that a standardized couch embedded with
machine-based fiducial device should be used in all imaging systems for
patients who are undergoing radiation therapy.
To assess the registration error of this machine-based fiducial fusion
- 57 -
method, target registration error (TRE) were measured for different anatomic
structure landmarks surrounding the treatment area. Besides the
misalignment error of PET and CT image, the target registration error (TRE)
includes the localization error of the anatomical structure landmark, error
due to partial volume effect and internal organ movement between PET and
CT scan.
Two observations were notable. First, the target registration errors
of all methods exceed alignment uncertainties nominally associated with PET
and CT image fusion techniques. It was important to consider whether
alignment data is quoted for a single target centroid or considered multiple anatomical points outside the disease site. Manual alignment was expected to be superior for biological targets since the operator can sacrifice alignment of other anatomies in favor of the target of interest. Second, patients might be shifted or distorted between PET and CT scans despite mask immobilization. Manual and automatic methods are relative measurements and do not indicate their absolute location in relation to isocenter as do fiducial alignments. Unlike diagnostic radiological exams including PET/CT, both relative fusion error and isocenter registration errors should be considered for the accurate treatment of patients undergoing radiation therapy.
Non-rigid body deformable registration is considered a solution to
- 58 -
compensate patient’s shift and internal organ movement between CT and PET
scans. However, a Multi-Institution Deformable Registration Accuracy
Study (MIDRAS) showed a substantial inconsistency in reported shifts (64).
Another investigation also showed that the mean registration error resulted from deformable registration was slightly greater than that from manual registration for head and neck radiation therapy treatment planning (65).
Immobilization device is recommended both in CT and PET scans to minimize patient’s movement and improve treatment planning, delivery, and assessment.
2.6 Conclusion
The small fiducial registration errors demonstrated that the fiducial board can be used as the reference for patient treatment position relative to the treatment machine isocenter. The manual image fusion method is considered to be most accurate in this study. The evaluation of the registration errors for ten patients with head and neck cancer showed that except base of skull, there were no significant differences in the target registration errors of specified anatomical landmarks using manual or machine-based fiducial image fusion method. Using the fiducial board in both CT and PET imaging procedures enables to map patient’s diagnostic position to the treatment position. The CT-PET image fiducial fusion based
- 59 - on the fiducial board is invariant to the treatment machine isocenter, which can improve the efficiency of target localization and delineation in radiation therapy treatment planning.
- 60 -
Chapter 3 An Automatic Method for PET Target Segmentation Using a Lookup Table Based on Volume and Concentration Ratio
3.1 Abstract
Accurate evaluation of functionally significant target volumes in combination with anatomic imaging is of primary importance for effective radiation therapy treatment planning. In this study, a method for rapid and accurate PET image segmentation and volumetrics based on phantom measurements and independent of scanner calibration was developed. A series of spheres ranging in volume from 0.5 mL to 95 mL were imaged in an anthropomorphic phantom of human thorax using two commercial PET and
CT/PET scanners. The target to background radioactivity concentration
ratio ranged from 3:1 to 12:1 in 11 separate phantom scanning experiments.
The results confirmed that optimal segmentation thresholding depends on
target volume and radioactivity concentration ratio. This information can be
derived from a generalized pre-determined “lookup table” of volume and
contrast dependent threshold values instead of using fitted curves derived
from machine specific information. A three-step method based the PET image intensity information alone was used to delineate volumes of interest.
- 61 -
First, a mean intensity segmentation method was used to generate an initial estimate of target volume, and the radioactivity concentration ratio was computed by a family of recovery coefficient curves to compensate for the partial volume effect. Next, the appropriate threshold value was obtained from a phantom-generated threshold lookup table. Lastly, a threshold level set method was performed on the threshold value to further refine the target contour by reducing the limitation of global thresholding. The segmentation results were consistent for spheres greater than 2.5 mL which yielded volume average uncertainty of 11.2% in phantom studies. The results of segmented volumes were comparable to those determined by contrast-oriented method and iterative threshold method (ITM). In addition, the new volume segmentation method was applied clinically to ten patients undergoing
PET/CT volume analysis for radiation therapy treatment planning of solitary lung metastases. For these patients, the average PET segmented volumes were within 8.0% of the CT volumes and were highly dependent on the extension of functionally inactive tumor volume. In summary, the current method does not require fitted threshold curves or a priori knowledge of the
CT/MRI target volume. This threshold method can be universally applied to radiation therapy treatment planning with comparable accuracy, and may be useful in the rapid identification and assessment of plans containing multiple targets.
- 62 -
3.2 Introduction
Positron emission tomography (PET) using glucose analog
[18F]2-fluoro-D-2-deoxyglucose (FDG) was developed essentially as a diagnostic tool for neoplasms detection (66). FDG-PET has been used in staging, treatment response assessment and restaging after therapy, and prognosis for several neoplasms (67). In addition to the significant diagnostic role in oncology, FDG-PET is increasingly being used in target volume definition as a planning tool for radiotherapy (68). The role of
FDG-PET in radiation therapy treatment planning has been investigated for several malignancies, including lung (69), head and neck (70, 71), brain (72), cervix (73), and other tumor sites. FDG-PET imaging has had a great impact on gross tumor volume (GTV) definition, especially for lung cancer (74).
However, there is no apparent general agreement about a uniformly applicable method for accurate and automatic target volume delineation (68).
The accuracy of GTV definition is essential in conformal radiation
therapy such as intensity modulated radiation therapy (IMRT) (75).
Conventionally, GTV was based on volume data derived from CT scanning.
However, CT provides relatively low contrast for soft tissue, which makes it difficult to differentiate the malignancy when the tumor has similar density with normal tissue. Previous investigation suggests that FDG-PET has the
- 63 -
potential to provide more accurate GTV definition and reduce inter-observer
variability (76). Studies reported that FDG-PET based GTV definition is
superior to the GTV defined by CT alone for a moving target (77). However, compared to CT or MRI images, FDG-PET images have low spatial resolution,
high partial volume effect, and insufficient anatomical details that result in
difficulty defining the exact borders of many types of tumors. More attention
and efforts are necessary to incorporate PET functional imaging information
into radiation therapy treatment planning.
Currently, threshold-based approaches are the most widely used
automatic methods for PET target delineation in research and clinical
applications. Although other techniques have been reported recently with
limited success such as region-growing (78), statistical model based (79),
gradient-based (80), and PET-CT based (81) methods, they require further
clinical validation. Many recent investigations, including this report,
indicate that the use of fixed or absolute value thresholds have limited
accuracy, especially for small volume or low contrast targets (82, 83). The
optimal threshold depends upon the source/background image intensity ratio
(S/B ratio) and the target volume.
Accordingly, an adaptive thresholding method was developed using a
CT volume as the initial estimate of target volume (84). In this method, a family of exponential threshold-volume curves for different S/B ratios was
- 64 -
determined from the data fitting of an initial sphere phantom study.
Depending on measured target S/B ratio, the given CT volume was applied to
the corresponding curve to yield the desired threshold for target delineation.
Similarly, Daisne et al. (85) used a threshold vs. S/B ratio curve to obtain the
desired threshold. Based on the same hypothesis of threshold-volume-S/B
ratio relationship, an iterative threshold method was recently developed (86).
Instead of using CT volume as a prior knowledge, the threshold derived from fitted curves and thresholding segmentation procedures were performed iteratively until convergence to a unique value was reached. Another local contrast-based method employed similar iterative technique as well as slice-by-slice analysis to obtain the desired threshold for each slice instead of one global threshold for the whole volume (87, 88). The iterative methods do not need the CT volume for an initial estimate of the target volume (86, 88).
The adaptive thresholding method (84) worked sufficiently well if an accurate
CT target volume was available. However, the convergence of the iterative procedure is questionable when applied to a target volume smaller than 2.5 mL. Moreover, some methods apparently use the S/B ratio interchangeably with physical FDG radioactivity concentration ratio of target and background
(84). Unlike the above techniques, Black’s mean SUV method (89) and a contrast-oriented method (90) employ a linear relationship between threshold and mean target intensity iteratively to yield the desired threshold. All
- 65 -
these methods require a parameterized curve fitting that needs to be modified
to accommodate different scanners and reconstruction protocols.
Furthermore, using global threshold contouring may cause problem in target delineation for tumors with heterogeneous activity distribution.
In order to improve on current limitations of thresholding methods
and to provide a rapid and convenient method for clinic image segmentation
based on PET images alone, a combined multi-step thresholding and level set
segmentation method was developed for radiotherapy target delineation.
The experimental results on phantom and patient data studies demonstrate
this method to be robust and machine independent with adequate accuracy for
wide range of target volumes and S/B ratios.
3.3 Methods and materials
Phantom studies were performed to investigate and quantify the relationship between optimal threshold, target volume, and target-background radioactivity concentration ratio. A series of spheres with volumes ranging from 0.5 to 95 mL were injected with a uniform concentration of [18F]2-fluoro-D-2-deoxyglucose (FDG) and suspended in a
Jaszczak phantom which imitates the size and shape of human thorax. To
emulate the typical background radioactivity in patient studies, 0.145 μCi/mL
FDG was used as background radioactivity in the phantom. Although the
- 66 -
wall of each hollow sphere contains no radioactivity, the wall thickness was approximately 1 mm which is below the PET image resolution (4mm) and was not included in the reconstructed images. The spheres were dispersed within the phantom with at least 4 cm of space between adjacent sphere edges.
Although spatial resolution and sensitivity are degraded in the radial position away from the center due to parallax error, it has been reported that the radial position of spheres has no observable influence on the threshold required to obtain a correct target (87). The sphere volumes were measured by an infusion syringe pump (Graseby 3400, Graseby Medical Ltd.) with a precision of ±0.1 mL, which was cross calibrated with a graduated cylinder.
The FDG radioactivity concentrations were monitored with a dose calibrator
(Deluxe Isotope Calibrator II, Victoreen, Inc.).
One set of spheres (1.0, 2.0, 5.7, 8.3, 11.6, 18.9, 19.3, and 95.0 mL) was scanned by a Philips Allegro PET scanner. The target-to-background radioactivity concentration ratio of 3:1, 6:1, 8:1, and 12:1 were applied. Each scan and subsequent measurements were repeated twice. The Allegro scanner has an 82 cm detector ring diameter and an axial field of view (FOV) of 18 cm. The spatial resolution near the center is approximate 5.5 mm and
5.6 mm in the transverse and axial direction respectively (91). Another set of spheres (0.5, 1.2, 2.6, 5.6, 11.6, and 26.7 mL) was scanned by a Philips Gemini
TF PET/CT scanner with 90 cm detector ring diameter and 18 cm axial FOV.
- 67 -
The spatial resolution near the center is approximate 4.6 mm (18). The
concentration ratio of 3:1, 6:1, and 12:1 were used in each of these spheres and
single set of scan was taken for this data set. The imaging procedure used to
scan the phantoms utilized the default clinical protocol set up imaging lung
cancer for each scanner. This procedure consisted of a transmission scan
(Allegro) or a CT scout (Gemini TF) and then a scan to define the PET imaging
field of view, followed by an emission scan consisting of two fields at a scan
rate of 3 min/bed position. Upon scan completion, a 3D row-action
maximum-likelihood algorithm (3D-RAMLA) reconstruction was performed
for the Allegro scanner. For the Gemini TOF scanner, time of flight ordered
subsets expectation maximization (TOF-OSEM) was performed. Both scan
reconstructions incorporated decay correction, background subtraction, and attenuation correction as performed in our clinical protocols. The pixel size of
the reconstructed PET image was 4 × 4 mm2 with a slice thickness of 4 mm.
Reconstructed PET images were analyzed by a in-house developed Matlab program (The MathWorks, Inc., version 7.5 release 2007b). When the difference between the calculated target volume and actual physical volume was small (<0.02 mL), an optimal threshold was determined for each sphere using a binary search program. Instead of utilizing machine specific curve fitting parameters, a simple six group threshold lookup table, (Table 3.1) comprised only of target volumes and concentration ratios, was generated
- 68 - based on the phantom experiment results. The workflow of the PET image segmentation method is shown in Figure 3.1.
Step 1 1. Image input
Step 2 2. Mean intensity method
3b. S/B ratio
Step 3a-d 3a. Initial volume 3c. Recovery coefficient
3d. Concentration ratio
4. Derive the desired Step 4 threshold from a pre-determined lookup table
Step 5 5. Threshold level set method
Step 6 6. Contour and volume
- 69 -
Figure 3.1 Workflow of the current PET target segmentation method. First, a mean intensity method was used to obtain the initial volume and S/B ratio
(steps 1, 2, 3a, and 3b). Next (steps 3c and 3d), the concentration ratio was obtained by applying the recovery coefficient table (Table 3.2). Based on the
initial volume and the recovered concentration ratio, a desired threshold was
then selected from the threshold lookup table (step 4). Lastly, this threshold was used to perform a standard level set method to delineate and estimate the final target volume (steps 5 and 6).
The first step in the workflow diagram is to acquire the PET image and identify the target region of interest. A mean intensity segmentation method (step 2) was developed for the current work to derive the initial estimate of the target volume and concentration ratio. This method is based on the hypothesis that optimal threshold intensity has a proportional linear relationship to mean target intensity and mean background intensity. The
current phantom study shows the coefficient of determination (R2) of this
linear relationship was 0.94. Equation 1 describes this linear relationship:
Threshold intensity = 0.66× mean target intensity (3.1) + 0.30× mean background intensity
- 70 -
This equation was derived by multiple linear regression model data
fitting, using the measurements obtained by the current phantom study.
Any pixel within predefined ROI whose intensity is below 40% of the
maximum was considered as background. The mean background intensity
was defined as the average of background pixel intensity. In the mean
intensity method, an iterative procedure was designed to utilize this equation
to obtain the initial volume estimate. The first segmentation was acquired using 40% threshold value (84). The initial mean target intensity and mean background intensity were measured to calculate the new threshold intensity according to equation 3.1. Per each iteration, a new segmentation, new mean target, and background intensity were computed based on the proceeding threshold intensity and this result was used to calculate the successive threshold intensity. By repeating this procedure, the value of the threshold converged to a fixed amount within 15 iterations. The convergent threshold value was used to obtain the initial estimate of target volume (step 3a, Figure
3.1).
The output of mean intensity segmentation also yielded the S/B ratio
(step 3b, Figure 3.1) which is the ratio of the maximum image intensity of the target to the average background intensity excluding the target. The S/B ratio is associated with the physical radioactivity concentration ratio but degraded by partial volume effect (PVE) due to finite spatial resolution (92).
- 71 -
In this study, the radioactivity concentration ratio was recovered (step 3d and
3c, Figure 3.1) by a family of recovery coefficient curves (Table 3.2) to compensate the partial volume effect, especially for small volumes, based on the initial estimate of target volume and measured S/B ratio. In step 4, the recovered radioactivity concentration ratio combined with the initial estimate of target volume provided sufficient information to find a desired threshold from the threshold lookup table (Table 3.1).
Next, this threshold was used to perform a standard threshold level set method (93) to delineate the target for each slice (step 5). Using the threshold level set method, the contour will be smoothed and pixel leakage due to global thresholding can be prevented. This method was implemented using a well-known and utilized threshold level set module available as freeware on the internet (Insight Segmentation and Registration Toolkit,
Kitware, Inc., supported by National Library of Medicine, version 3.6). The theoretical foundation and application to medical image for this program is discussed in detail by Sethian et al. (93). We experimentally selected the program-specific weighting factor between the normal direction propagation and contour curvature such that clinically relevant smoothness of the contour is maintained while preserving the size and shape information. As the final step after contour smoothing (step 6), the target volume was calculated based on the sum of segmented target area for each slice multiplied by the slice
- 72 -
thickness.
The current method of PET image segmentation was used to compute
volumes for the sphere images from all experiments using two different
scanners. These images were also analyzed to derive the optimal thresholds
and used to generate the lookup table (Table 3.1). As a self-consistency check,
these volumes were compared to the physical volume measured for each
sphere (Table 3.3 and Table 3.4). For comparison, we recalculated the
volumes of spheres using the contrast-oriented method (90) and the iterative threshold method (ITM) (88). The current method was also applied to the image data from ten patients with solitary lung nodules who were undergoing volume analysis for radiation therapy treatment planning. From this clinical study, preliminary validation of the current image segmentation method was obtained by comparing the CT segmented lung nodule volumes with the calculated volume.
3.4 Results
The foundation of the present work is the relationship among optimal threshold, target volume and radioactivity concentration ratio. The optimal
thresholds for each sphere volume in all experiments and radioactivity
concentration ratio combination are plotted in Figure 3.2.
- 73 -
Figure 3.2 Optimal thresholds yielding correct target volume for each sphere target with different volume and radioactivity concentration ratio in Allegro scanner (a) and Gemini TF scanner (b).
- 74 -
The optimal thresholds were calculated based on the measured target and background activity. In both the Allegro and Gemini TF scanners, the minimal average background activity of all experiments was 2280 counts/pixel and the target activity was always higher than background activity for all spheres. Therefore, the maximum associated statistical noise was 2.1% of a single pixel value. In Figure 3.2(a), each data point was the average of two
repeated measurements for the Allegro scanner. The average standard
deviation of for all points was 2.8% in absolute threshold. In Figure 3.2(b),
all the data points were from a single scan in Gemini scanner. Each optimal
threshold was shown to have a complex dependence on sphere volume and radioactivity concentration ratio, therefore, no simple curve fit could model the data. To summarize the relationship among optimal threshold, target volume and radioactivity concentration ratio, a lookup table was generated.
Based on the similarity with neighboring measured values, the entire set of
measured optimal thresholds was categorized into six similar groups (Table
3.1). These groups were associated with two classes of target volumes and
three classes of concentration ratios. The separation of these groups was
determined by minimizing the summation of each group’s deviation. Each
cell in Table 3.1 is the numerical average of at least four measurements of
optimal threshold expressed as a percent of the maximum intensity and
corresponds directly to the measured volume and concentration ratio for each
- 75 -
group.
Table 3.1 Threshold lookup table consists of target volume and
radioactivity concentration ratio. A desired threshold for target
segmentation is chosen from this table based on the initial estimate of target
volume and concentration ratio.
C: Concentration ratio
V: Volume C < 4:1 4:1 ≤ C < 10:1 C ≥ 10:1
V < 2.5 mL 74.9% 59.2% 54.3%
V ≥ 2.5 mL 54.7% 44.4% 39.7%
With respect to the mean intensity method (step 2, Figure 3.1), the
phantom data analysis shows that the optimal threshold has a direct
proportional linear relationship with the mean target intensity and mean
background intensity. A multiple linear regression model established this relationship as equation 3.1. Although the image pixel intensity units in this phantom study were counts/pixel, this equation could be adapted for other units such as activity/mL or SUV since no absolute offset appears in equation
3.1. All the variables expressed in intensity units can be normalized to a
percentage of target maximum intensity. The coefficient of determination
- 76 -
(R2) of multiple linear regression curve fitting was 0.94. The relative
deviation of threshold calculation from measured optimal thresholds was
within ±15% for all data points.
The output of mean intensity segmentation also includes the S/B ratio.
From the sphere phantom data, it was noted that the S/B ratio is not equivalent with the target-to-background radioactivity concentration ratio, especially for small size spheres due to the partial volume effect. According to the relationship among the concentration ratio, sphere volume, and measured S/B ratio shown in Figure 3.3, the concentration ratio could be
substantially recovered from the initial estimate of target volume and S/B ratio.
- 77 -
Figure 3.3 Plots of S/B ratio versus measured sphere volume for the phantom scanned in Allegro (a) and Gemini TF scanner (b).
- 78 -
Similarly, because no simple curve can fit all the data points, as
indicated in Figure 3.1 step 3c, a recovery coefficient table (Table 3.2) was constructed according to the relationship shown in Figure 3.3 for partial volume effect compensation. The volume division in Table 3.2 was chosen based on the inflexion point on S/B ratio-volume curves (see Figure 3.3).
Based on Table 3.2, with the initial estimate of target volume and measured
S/B ratio, the concentration ratio is recovered.
Table 3.2 Recovery coefficient table. Concentration ratio (C) was corrected and recovered for partial volume effect based on initial estimate of target volume and measured source/background (S/B) ratio.
SBR: V: Initial Volume
source/background ratio V ≤ 4 mL V > 4 mL
2.0 ≤ SBR < 3.5 C = 0.4 (7-V) SBR C = 1.15 SBR
3.5 ≤ SBR < 5.0 C = 0.6 (7-V) SBR C = 1.33 SBR
5.0 ≤ SBR < 6.5 C = 0.6 (7-V) SBR C = 1.40 SBR
SBR ≥ 6.5 C = 0.6 (7-V) SBR C = 1.60 SBR
- 79 -
The recovered concentration ratio combined with the initial estimate of target volume provided sufficient information to derive the desired threshold from the threshold lookup table (Table 3.1) to delineate the target volume (step 4 Figure 3.1). Table 3.3 and Table 3.4 summarize the volume calculation results using the current method for each sphere imaged in the
Philips Allegro PET scanner and Gemini TF PET/CT scanner respectively.
Each row in Table 3.3 and Table 3.4 shows the average and standard deviation in volume estimation using all scans of each sphere for multiple concentration ratios. For spheres greater than 2.5 mL, the average volume estimation error was measured to be 11.2%. Since the same threshold was applied to different spheres within the same group, it is noted that sphere No.1 is always overestimated and sphere No.2 is always underestimated.
- 80 -
Table 3.3 Volume calculation results using current method for spheres were imaged in Philips Allegro PET scanner. Each sphere was scanned with 4 different concentration ratios ranging from 3:1 to 12:1 in two-fold redundancy.
Actual Volume Average STD Sphere No. (mL) estimate (mL) (mL)
1 1.0 1.1 0.4 2 2.0 1.8 0.6 3 5.7 6.4 2.7 4 8.3 7.8 0.7 5 11.6 11.7 1.3 6 18.9 17.9 2.3 7 19.3 19.1 1.8 8 95.0 93.8 4.7
- 81 -
Table 3.4 Volume calculation results using current method for spheres were
imaged in Philips Gemini TF PET/CT scanner. Each sphere was scanned
with 3 different concentration ratios ranging from 3:1 to 12:1.
Actual Volume Average STD Sphere No. (mL) estimate (mL) (mL)
9 0.5 1.2 1.2 10 1.2 0.5 0.3 11 2.6 1.8 0.4 12 5.6 4.7 0.9 13 11.6 10.8 1.2 14 26.7 25.4 2.6
The results of the comparison with the recalculated sphere volumes using the contrast-oriented method (90) and the iterative threshold method
(ITM) (88) are shown in Table 3.5. For spheres No.1, 2, 9 and 10 (1 mL, 2 mL,
0.5 mL and 1.2 mL respectively), the iterative threshold method did not converge and failed to yield proper segmentation results. Hence, these spheres were excluded from Table 3.5.
- 82 -
Table 3.5 Volume estimation uncertainty (% error) comparison of the current method with other published methods (88, 90).
Contrast-oriented Iterative threshold Present Sphere volume (V) method method method
Small (2.5mL Medium (4mL≤V<12mL) 31.2% 14.6% 13.7% Large (V≥12mL) 17.8% 6.5% 6.2% Average 25.7% 12.5% 11.2% *Results of sphere No.1, 2, 9 and 10 (1 mL, 2 mL, 0.5 mL and 1.2 mL respectively) were not included since the iterative threshold method failed for these spheres due to non-convergence problem. In summary, all methods yielded comparable results for the sphere targets. The average volume estimation error for small spheres using the current method is 30.6% compared to 41.9% and 45.9% for the contrast-oriented method and the iterative threshold method respectively. The overall average calculation errors for the contrast-oriented method, the iterative threshold method and the current method are 25.7%, 12.5% and 11.2% respectively. Ten image volumes from patients with solitary lung nodules were processed - 83 - with the current method. The calculated volumes are listed below in Table 3.6 and are compared to the GTVs determined by CT volumes. The average error of the volume estimation is 8.0%. As an illustration, the axial CT slice and the co-registered PET slice of lesion No. 2 are shown in Figure 3.4. Based on the initial estimate of volume (>2.5 mL) and recovered concentration ratio (> 10:1), the threshold value of 39.7% in the lookup table (Table 3.1) was used. The manually defined GTV on CT image is shown on the left panel. The target delineation obtained by the current method on PET image is shown on the right panel. Figure 3.4 Comparison of CT volume and the segmented PET volume for lesion No. 2. The left panel is the axial CT slice with the manually defined GTV (green). The right panel is the co-registered PET slice with the automatically delineated contour as applied by the current method (pink). - 84 - Table 3.6 Target volume calculation for clinical patient images compared with the CT-defined GTV. GTV defined by CT PET volume with Lesion No. Difference volume (mL) proposed method (mL) 1 2.9 2.7 -6.9% 2 6.0 6.1 1.7% 3 23.7 22.3 -5.9% 4 1.2 1.1 -8.3% 5 18.3 13.9 -24.0% 6 5.3 4.2 -20.8% 7 64.2 66.2 3.1% 8 2.6 2.5 -3.8% 9 1.1 1.1 0% 10 7.9 7.5 -5.1% Average 8.0% - 85 - 3.5 Discussion In this study, an automatic PET target delineation method was developed and evaluated with sphere phantom and patient data. This method is based on a phantom-based threshold lookup table and a threshold level set function technique, which is independent of the scanner calibration factor and does not require any fitted threshold curve or a priori information of target volume as with previous studies (84, 86, 88, 90). The phantom studies showed that the optimal threshold yielding correct target delineation depends on the target volume and target to background concentration ratio as outlined in Figure 3.1. It is noted that the fitted optimal threshold vs. sphere volume curves were essentially similar in appearance with exponential decay function. However, simple curve fitting of this data (84, 86) has limited practical benefit for prospectively using a continuous spectrum of optimal thresholds corresponding volumes. The targets with similar volume and same concentration ratio may have substantial different optimal threshold because of the inherent noise associated with PET images, uncertainty in absolute volume or radioactivity concentration measurement, and/or variations in the reconstruction algorithms used by different scanners (e.g., sphere No. 6 and 7 in Figure 3.2a). The repeated experiments using a different phantom and scanner showed much less fluctuation in the threshold values (Figure 3.2b). Nevertheless, - 86 - the average of fluctuation portion in Figure 3.2a is reasonably consistent (±5%) with the threshold value derived from Figure 3.2b. In addition, one may observe that these curves reach a local minimum when the sphere volume ranges between 2 to 10 mL yielding the same threshold for multiple volumes. Similar results are also noted in the work published by Drever et al. (88), but are not specifically cited here. For small volume target (< 2.5 mL), the precise input of target volume for the curve fitting method is required since a small change in this input can result in a large change in the resulting threshold. Furthermore, each fitted curve represents a unique and singular concentration ratio. Hence, interpolation values between fitted curves are required to arrive at the appropriate threshold value. Instead of using fitted curves as a basis (84, 86), a lookup table (Table 3.1) was used to derive the desired threshold in the current method. The apparent advantage of the lookup table scheme is that the threshold value in each group is relatively insensitive to the properties of the imaging device or the reconstruction method used. The threshold lookup table only requires a first-order estimate of target volumes and corresponding concentration ratios from a PET image. The relative stability to the initial estimate used in the current method shows a significant advantage compared to previous investigations (84, 86, 90). As long as these initial estimates fall into the correct group or category, the threshold derived from the lookup table method - 87 - will result in acceptable target delineation. Only 6 data groups of target size and concentration ratio were assigned in the threshold lookup table. The number of groups and the lookup table structure were limited based on the observer’s ability to judge target volume from a PET image. Without precise estimates of target volume and concentration ratio, an increase in the number of data categories would not improve the accuracy of target segmentation. The boundary points were determined so that most of the initial estimated volumes and the recovered concentration ratios of all sphere target fell into the correct group. The determination of these boundary points is a first-order estimate and may be the subject of further experimentation and refinement. No further attempt was made to optimize the number of data categories, range or boundary values of each group for this work, which may be a focus of future studies. The current study showed that using a single threshold in the lookup table within a defined range of volume and concentration ratio will result in acceptable target delineation. In current practice, PET image display is most commonly presented in unit of SUV, although other units (e.g., relative specific activity) may be used (94). Since the present work relies on the input relative intensity values, the segmentation results are unaffected by the choice of unit used for the image intensity display (cts/pixel, µCi/mL or SUV), thereby eliminating the influence of any machine calibration factors (e.g., counts/MBq) on the segmentation - 88 - results. There are different ways to obtain the initial estimate of target volume. The volume manually determined from CT images can serve as a source of the initial estimate of target volume (84). However, this volume is not always available. Jentzen, et al. (86) have shown that a volume can be derived by using an iterative procedure to locate a fixed point on the fitted curve. However, when applied to our phantom data, this iterative threshold method may not provide a unique or valid solution because of the non-convergence problem. In this present work, a mean intensity method is used to obtain an initial estimated target volume (Figure 3.1). By incorporating the mean background intensity and omitting any linear offset, this method can use image intensity with arbitrary units normalized to the maximum intensity. Furthermore, this method provided suitable estimate well for larger target volumes (±10%), but the volume estimation error for targets smaller than 2.5 mL was considerable (greater than 100%). Therefore, this mean intensity method was only used as an approximate initial estimate of target volume. Similar to the iterative threshold method (86), in order to calculate the mean intensity of target and background, an accurate volume segmentation was necessary. To solve this problem, a separate iterative routine was developed. In the current study, the iteration of the threshold calculation always leads to a fixed convergent point for all of - 89 - the phantom data. This is an advantage compared to the threshold-volume exponential curve fitting methods (86, 88), where a small deviation in the estimated target volume results in a substantial change in the calculated threshold and produces a large change in the successive volume estimation. This in turn magnifies the error which causes the iteration not to converge, especially for target volumes smaller than 2.5 mL. In addition to the target volume, the optimal threshold also depends on the concentration ratio which is one of two variables in the threshold lookup table. The concentration ratio is defined as the FDG radioactivity concentration ratio between the target and background region, which cannot be directly derived from the image because of partial volume effect. The concentration ratio is related to the S/B ratio but may not be directly proportional to the S/B ratio. The S/B ratio can be directly determined from the image as long as the target volume is accurately delineated. To correct for the partial volume effect, a recovery coefficient table (Table 3.2) is used to recover the concentration ratio based on measured S/B ratio and initial estimate of target volume. With the initial estimate of target volume and recovered S/B ratio, an optimal threshold can be derived from the pre-determined lookup table. However, global thresholding using a single threshold for the entire image is problematic since only the intensity is considered (95). The spatial - 90 - information is overlooked and there is no guarantee that the pixels identified by the thresholding process are contiguous. In addition, simple thresholding can easily include extraneous pixels outside of the target or can miss isolated pixels within the target especially in a noisy environment such as a PET scan. In this present work, a level set method was applied using the derived threshold from the lookup table as guidance in order to improve the limitations due to global thresholding. This implicit interface evolving technique moves the contour to the boundary of the target and controls the curvature smoothness simultaneously (96). The constrained smoothness of the contour helps to avoid some of the pixel leakage. Furthermore, the level set method can decrease the sensitivity to the image noise. The sphere phantom volume estimation results in Table 3.3 and Table 3.4 show a uniform and consistent performance of the current method over a wide range of target volumes and concentration ratios in two different scanners. Table 3.5 shows the current method yields comparable results for sphere phantoms compared to the contrast-oriented method and the iterative threshold method (ITM). Extracranial lesions of less than 2 mL are usually included in an expanded region of irradiation, instead of being irradiated as a single focal target (e.g., head and neck nodal irradiation) (69, 70). Excluding these small spheres less than 2.5 mL, the average volume estimation error of the current method is 11.2%, which is comparable to other published results - 91 - (84, 86, 88, 89). However, the current method does not require any fitted threshold curve or a priori information of target volume. The parameters of the current method are applicable to small targets as well. Volume estimates using the current method for the targets less than 2.5 mL have larger associated uncertainties (50-100%). However, due to non-convergence problem, the curve fitting based methods (84, 86, 88) may fail or induce an error greater than 100% for these targets based on the current phantom study. Methods for PET image segmentation based on region-growing (78), gradient (80) or statistical model (79) have showed to be useful but require sophisticated model estimation, high computation cost and further clinical validation. An additional concern is the intrinsic low image resolution which affects the accuracy of PET target delineation. The PET image slice thickness was 4 mm and the image resolution was 4 mm by 4 mm in this study and the image voxel size was 0.064 mL. For the smallest sphere target in our phantom study with a volume of 1 mL, a segmentation inaccuracy of one voxel will translate into a 6.4% volume estimation error. To decrease the uncertainty below one voxel requires sub-pixel interpolation. In current study, a bilinear interpolation method was used to resize each slice to 1 mm by 1 mm. To reduce the computation cost, no interpolations between slices in superior/inferior directions were performed. By this means, 1 voxel offset - 92 - only results in 0.4% volume estimation error. Using the same threshold to segment the original images and interpolated images resulted in a discrepancy of at least 4% in the volume estimation. We have also noted that the fitted optimal threshold vs. sphere volume curves were essentially similar in appearance with exponential decay function. However, one may observe that these curves reach a local minimum when the sphere volume ranged between 2 to 10 mL. Similar results are also noted in the work published by Drever et al (88), but were not specifically cited. Computer simulation of model based generated sphere targets has been run accounting for scanner spatial resolution and target volume variance which may provide a theoretical basis for this observation. These modeling simulation are ongoing and validated theoretical basis is under investigation. To demonstrate the usefulness in clinical practice, the current method was applied to patients with lung tumors and the volumes were compared to the physician defined GTV in CT images. The average volume estimate error of 8.0% for 10 lesions with a volume ranging from 1.1 mL to 64.2 mL is comparable to the clinical results in other investigations (80, 86, 90). These lesions were isolated from normal lung tissue and clearly delineable in CT images, therefore the observer variation in GTV definition is reduced. It is noted that 8 out of 10 PET volumes were smaller than the corresponding CT volumes. This is consistent with reports that show the functional active - 93 - volume observed in PET versus CT anatomic imaging is smaller due to inclusion of non-functional tissues in the CT images (82, 97). In particular, by visual inspection, the larger estimation errors (> 20%) associated with lesion No. 5 and 6 were due to the overestimate of the CT-defined GTV compared to the functional active volume shown in corresponding PET image. On the other hand, indications of patient respiration movement during PET scanning were found in the cases where the calculated PET volume was greater than CT volume. Previous investigations also point out that these respiration movements blur and enlarge the outlined target volume for PET images (77, 80). 3.6 Conclusion This study showed that the optimal threshold for FDG-PET target delineation depends on the target volume and the target-to-background radioactivity concentration ratio which can be obtained from image data alone without the need for absolute machine activity quantification. The relationship between the optimal threshold, target volume, and concentration ratio is summarized in a lookup table (Table 3.1). The lookup table scheme precludes the complication of lack of convergence that may occur with the application of iterative methods. This study further shows that mean intensity is an appropriate method to obtain the initial estimate of target - 94 - volume. The radioactivity concentration ratio can be recovered based on the initial target volume and measured target S/B ratio. The current method results in an accurate, clinically useful PET image segmentation methodology with acceptable target delineation as shown in the phantom studies and clinical data. The lookup table of desired thresholds provides guidance for clinical PET target delineation for a wide range of target sizes and radioactivity concentration ratios encompassing the full range of FDG uptake values observed clinically. This method facilitates accurate target delineation and GTV definition for radiation therapy treatment planning, and provides a useful tool to facilitate dose escalation to the viable tumor volume. - 95 - Chapter 4 Conclusions and Suggestions of Future Work 4.1 Conclusions The combined methodologies of the machine-base CT-PET fiducial fusion introduced in Chapter 2 and the new thresholding PET target segmentation introduced in Chapter 3 results in an overall improvement in the effective delivery of therapeutic radiation to patients. A machined-based fiducial system was developed to map the patient’s diagnostic position to treatment position. The small fiducial registration error demonstrated that the fiducial board can be used as the reference of treatment machine isocenter. The evaluation of the registration errors for patients with head and neck cancer showed that except base of skull, there were no significant differences in the target registration errors of specified anatomical landmarks between machine-based fiducial image fusion method and manual method which is considered to be most accurate in this research. This registration accuracy difference for base of skull was due to patient’s movement relative to the fiducial board between different scans and the uncertainty in anatomical landmark localization. Therefore, the machine-based fiducial system assists in CT-PET image fusion with shorter processing time and acceptable accuracy, - 96 - which provides improvement for the target localization and delineation for radiation therapy treatment planning. In addition, the results of the phantom studies and the clinical applications in Chapter 3 have shown that the new thresholding technique results in an accurate, clinically useful PET image segmentation methodology with acceptable target delineation. The lookup table of desired thresholds provides guidance for clinical PET target delineation for a wide range of target sizes and radioactivity concentration ratios encompassing the full range of FDG uptake values observed clinically. Accurately localized and delineated tumor volume is expected to decrease the conformality index of the tumor dose (tumor volume/target volume) and spare the normal tissue more efficiently. The new PET image segmentation method is also expected to be independent of scanner calibration factors. 4.2 Future studies Future work is necessary for further validation of the proposed CT-PET image fiducial fusion and PET image segmentation methods. Application of the current methods to other disease sites and treatment configurations will be helpful to translate the current research into the improvement of clinical outcome and treatment safety. The suggested future work includes: 1) further validation of CT-PET image fusion accuracy; 2) application of the fiducial on treatment couch for patient localization during - 97 - daily treatment; 3) validation of the machine independence of the new PET image segmentation method on different PET scanners; 4) application of the automatic PET segmentation results (GTV modification) for clinical outcome improvement. 4.2.1 Validation of CT-PET image fusion Further validation of the machine-based CT-PET fiducial image fusion needs to be continued. The challenge of the image fusion validation is that there is no appropriate objective measurement for registration error when applied in the clinic. The registration error derived from phantom study usually does not include patient’s shift, internal organs movement and deformation between different scans. External fiducials placed on patient’s skin or internal fiducials placed inside of patient may be considered as objective references in image fusion to evaluate the registration error. However, it is still subject to the limitations such as fiducial localization error (FLE), fiducial misplacement between different scans, internal organ movement and deformation. Moreover, placing fiducial markers on patient time consuming and may not be always available in clinic, especially for internal fiducials. Measurement of target registration error for specific anatomical landmarks by multiple experienced specialists is expected to reduce the uncertainty of the registration error evaluation. However, these - 98 - measurements are based on visual inspection, which is subjective and limited by the observers’ knowledge and understanding of imaging mechanism and human anatomy. 4.2.2 Clinical application of the fiducial device in daily radiation therapy patient positioning With different immobilization equipments, e.g. stereotactic frame or alpha cradle whole body mould, the machine-based CT-PET fiducial fusion method can be used in stereotactic radiation therapy (SRT) or radiation therapies for other treatment sites such as lung and abdomen. These applications require further investigation and validation. In future work, the fiducial device can be used on treatment machine couch to assist in patient positioning and position verification in daily radiation treatment. The machine-based external line markers are reliable reference for the targets within patient’s treatment site. Aligning with the fiducial markers will speed up the patient positioning procedure and improve the accuracy of radiation therapy. An enhanced immobilization of the patient in the correct treatment position and subsequent shift or rotation of target isocenter of PTV is expected to produce the accuracy close to that provided by Gamma Knife treatment unit (0.1 mm). The phase II of this study will be validation of the daily patient positioning uncertainty using the machine-based fiducial device. - 99 - After fiducial-based alignment, the patient’s position can be verified by on-board imaging system and compared with planning CT. In recent literatures, many non-rigid body image registration methods were developed to improve the accuracy of image registration and correct for soft-tissue deformation during imaging (98). However, at least one investigation showed that the registration error resulted from global non-rigid body image registration is slightly greater than that achieved by manual registration (65). Another major concern is about the tumor volume geometry which is also deformed with other normal tissues in non-rigid body image registration. If the greater deformation is required, especially when the primary image and secondary image have a large discrepancy in patient’s position and the relative spatial relationship of one organ to another, the deformation turns into distortion of the fusion image. This kind of image distortion can impair the tumor definition in radiation therapy treatment planning. 4.2.3 Validation of the machine independence of the new PET image segmentation method In addition, the current research only demonstrated the application of the new thresholding PET target delineation in two Philips PET scanners (Allegro and Gemini TF), although the new method is expected to be - 100 - independent of scanner calibration factors. Further independent validation of the new thresholding method is necessary using different phantoms, different PET scanners, and clinical data as well. 4.2.4 Application of the automatic PET segmentation results (GTV adjustment) for clinical outcome improvement Several Investigations showed that GTVPET (gross tumor volume defined by PET) is usually smaller than GTVCT (gross tumor volume defined by CT only) for head and neck cancer (99-101). However, the PTV reduction based on the smaller GTVPET may result in suboptimal tumor control (102), though a smaller PTV can spare the normal-tissue more effectively (103, 104). Instead of shrinking the PTV, it is suggested to perform dose escalation only at the overlapping region of GTVPET and GTVCT, while keeping the PTV same as what is based on GTVCT. Assuming that the PET volume can be delineated accurately, the escalated dose is expected to result in improved tumor control, since PET imaging has high sensitivity and specificity for head and neck cancer (105) and the local tumor control is related to the delivered dose. Concurrently, the normal tissues are spared in same way as regular treatment plan due to the unchanged PTV. In order to demonstrate the clinical utility of this dose escalation scheme, further clinical trials would have to be performed and the related toxicity data must be obtained to show - 101 - the efficacy and the safety contributed by this dose escalation scheme. 4.3 Summary In this dissertation, it was proposed to develop: 1) a machine-based CT-PET fiducial fusion method for radiation therapy treatment planning and 2) a new thresholding PET image segmentation method which uses a threshold lookup table to assist in target delineation for image-guided therapy. In Chapter 2, we described how to fabricate a fiducial device and used this device to fuse PET and CT data sets from ten head and neck carcinoma patients resulting in a machine-based method for target localization and potential treatment with acceptable clinical accuracy. In Chapter 3, we developed a new thresholding PET image segmentation method using a threshold lookup table to perform accurate target delineation. Based on phantom and preliminary clinical studies, this method was shown to be clinically useful for PET target image segmentation and yielded comparable or superior accuracy to other reported methods. These combined methodologies are expected to result in an overall improvement in the effective delivery of therapeutic radiation to patients who are undergoing treatment for their disease. Based on these initial findings, it is important that further validation using different PET scanners is performed to confirm that the thresholding - 102 - technique is machine independent. Furthermore, it is equally important that the fiducial board be used during the actual treatment delivery with patients undergoing radiation therapy. If these future studies are accomplished, the translational aspect of bringing these treatment delivery concepts into the clinic will be potentially achieved. - 103 - Bibliography 1. Nutt R. 1999 ICP Distinguished Scientist Award. The history of positron emission tomography. Mol Imaging Biol. 2002 Jan;4(1):11-26. 2. Wagner HN, Jr. A brief history of positron emission tomography (PET). Semin Nucl Med. 1998 Jul;28(3):213-20. 3. Moses WW. Recent Advances and Future Advances in Time-of-Flight PET. Nucl Instrum Methods Phys Res A. 2007 Oct 1;580(2):919-24. 4. Budinger TF. PET instrumentation: what are the limits? Semin Nucl Med. 1998 Jul;28(3):247-67. 5. Votaw JR. The AAPM/RSNA physics tutorial for residents. Physics of PET. Radiographics. 1995 Sep;15(5):1179-90. 6. Rahmim A, Zaidi H. PET versus SPECT: strengths, limitations and challenges. Nucl Med Commun. 2008 Mar;29(3):193-207. 7. Karp JS, Surti S, Daube-Witherspoon ME, Muehllehner G. Benefit of time-of-flight in PET: experimental and clinical results. J Nucl Med. 2008 Mar;49(3):462-70. 8. Kinahan PE, Townsend DW, Beyer T, Sashin D. Attenuation correction for a combined 3D PET/CT scanner. Med Phys. 1998 Oct;25(10):2046-53. 9. Watanuki S. [Attenuation correction methods in PET and PET/CT: basic - 104 - principles and validity]. Nippon Hoshasen Gijutsu Gakkai Zasshi. 2006 Jun 20;62(6):797-803. 10. Nuyts J, Stroobants S, editors. Reduction of attenuation correction artifacts in PET-CT. Nuclear Science Symposium Conference Record, 2005 IEEE; 2005 23-29 Oct. 2005. 11. Habib Z, Marie-Louise M. Scatter Compensation Techniques in PET. 2007;2(2):219-34. 12. Zaidi H. Scatter modelling and correction strategies in fully 3-D PET. Nucl Med Commun. 2001 Nov;22(11):1181-4. 13. Brasse D, Kinahan PE, Lartizien C, Comtat C, Casey M, Michel C. Correction methods for random coincidences in fully 3D whole-body PET: impact on data and image quality. J Nucl Med. 2005 May;46(5):859-67. 14. Badawi RD, Lodge MA, Marsden PK. Algorithms for calculating detector efficiency normalization coefficients for true coincidences in 3D PET. Phys Med Biol. 1998 Jan;43(1):189-205. 15. Oakes TR, Sossi V, Ruth TJ. Normalization for 3D PET with a low-scatter planar source and measured geometric factors. Phys Med Biol. 1998 Apr;43(4):961-72. 16. Mawlawi O, Townsend DW. Multimodality imaging: an update on PET/CT technology. Eur J Nucl Med Mol Imaging. 2009 Mar;36 Suppl 1:S15-29. 17. Ford EC, Kinahan PE, Hanlon L, Alessio A, Rajendran J, Schwartz DL, Phillips M. Tumor delineation using PET in head and neck cancers: threshold contouring - 105 - and lesion volumes. Med Phys. 2006 Nov;33(11):4280-8. 18. Surti S, Kuhn A, Werner ME, Perkins AE, Kolthammer J, Karp JS. Performance of Philips Gemini TF PET/CT scanner with special consideration for its time-of-flight imaging capabilities. J Nucl Med. 2007 Mar;48(3):471-80. 19. Hoffman EJ, Huang SC, Plummer D, Phelps ME. Quantitation in positron emission computed tomography: 6. effect of nonuniform resolution. J Comput Assist Tomogr. 1982 Oct;6(5):987-99. 20. Papathanassiou D, Bruna-Muraille C, Liehn JC, Nguyen TD, Cure H. Positron Emission Tomography in oncology: present and future of PET and PET/CT. Crit Rev Oncol Hematol. 2009 Dec;72(3):239-54. 21. Zafra M, Ayala F, Gonzalez-Billalabeitia E, Vicente E, Gonzalez-Cabezas P, Garcia T, Macias JA, Vicente V. Impact of whole-body 18F-FDG PET on diagnostic and therapeutic management of Medical Oncology patients. Eur J Cancer. 2008 Aug;44(12):1678-83. 22. Gambhir SS. Molecular imaging of cancer with positron emission tomography. Nat Rev Cancer. 2002 Sep;2(9):683-93. 23. Le Bars D. Fluorine-18 and medical imaging: Radiopharmaceuticals for positron emission tomography. Journal of Fluorine Chemistry. 2006;127(11):1488-93. 24. Brock CS, Meikle SR, Price P. Does fluorine-18 fluorodeoxyglucose metabolic imaging of tumours benefit oncology? Eur J Nucl Med. 1997 Jun;24(6):691-705. 25. O'Doherty MJ, Macdonald EA, Barrington SF, Mikhaeel NG, Schey S. Positron - 106 - emission tomography in the management of lymphomas. Clin Oncol (R Coll Radiol). 2002 Oct;14(5):415-26. 26. Lardinois D, Weder W, Hany TF, Kamel EM, Korom S, Seifert B, von Schulthess GK, Steinert HC. Staging of non-small-cell lung cancer with integrated positron-emission tomography and computed tomography. N Engl J Med. 2003 Jun 19;348(25):2500-7. 27. Sasaki R, Komaki R, Macapinlac H, Erasmus J, Allen P, Forster K, Putnam JB, Herbst RS, Moran CA, Podoloff DA, Roth JA, Cox JD. [18F]fluorodeoxyglucose uptake by positron emission tomography predicts outcome of non-small-cell lung cancer. J Clin Oncol. 2005 Feb 20;23(6):1136-43. 28. Wong RJ, Lin DT, Schoder H, Patel SG, Gonen M, Wolden S, Pfister DG, Shah JP, Larson SM, Kraus DH. Diagnostic and prognostic value of [(18)F]fluorodeoxyglucose positron emission tomography for recurrent head and neck squamous cell carcinoma. J Clin Oncol. 2002 Oct 15;20(20):4199-208. 29. Grunwald F, Kalicke T, Feine U, Lietzenmayer R, Scheidhauer K, Dietlein M, Schober O, Lerch H, Brandt-Mainz K, Burchert W, Hiltermann G, Cremerius U, Biersack HJ. Fluorine-18 fluorodeoxyglucose positron emission tomography in thyroid cancer: results of a multicentre study. Eur J Nucl Med. 1999 Dec;26(12):1547-52. 30. Avril N, Rose CA, Schelling M, Dose J, Kuhn W, Bense S, Weber W, Ziegler S, Graeff H, Schwaiger M. Breast imaging with positron emission tomography and - 107 - fluorine-18 fluorodeoxyglucose: use and limitations. J Clin Oncol. 2000 Oct 15;18(20):3495-502. 31. Yoshida Y, Kurokawa T, Kawahara K, Tsuchida T, Okazawa H, Fujibayashi Y, Yonekura Y, Kotsuji F. Incremental benefits of FDG positron emission tomography over CT alone for the preoperative staging of ovarian cancer. AJR Am J Roentgenol. 2004 Jan;182(1):227-33. 32. Ioannidis JP, Lau J. 18F-FDG PET for the diagnosis and grading of soft-tissue sarcoma: a meta-analysis. J Nucl Med. 2003 May;44(5):717-24. 33. Price P, Jones T. Can positron emission tomography (PET) be used to detect subclinical response to cancer therapy? European Journal of Cancer. 1995;31(12):1924-7. 34. Apisarnthanarax S, Chougule P. Intravascular brachytherapy: a review of the current vascular biology. Am J Clin Oncol. 2003 Jun;26(3):e13-21. 35. Koukourakis G, Kelekis N, Armonis V, Kouloulias V. Brachytherapy for prostate cancer: a systematic review. Adv Urol. 2009:327945. 36. Lee NY, Terezakis SA. Intensity-modulated radiation therapy. J Surg Oncol. 2008 Jun 15;97(8):691-6. 37. Welsh JS, Patel RR, Ritter MA, Harari PM, Mackie TR, Mehta MP. Helical tomotherapy: an innovative technology and approach to radiation therapy. Technol Cancer Res Treat. 2002 Aug;1(4):311-6. 38. Harmon J, Van Ufflen D, Larue S. Assessment of a radiotherapy patient cranial - 108 - immobilization device using daily on-board kilovoltage imaging. Vet Radiol Ultrasound. 2009 Mar-Apr;50(2):230-4. 39. Rohrer Bley C, Blattmann H, Roos M, Sumova A, Kaser-Hotz B. Assessment of a radiotherapy patient immobilization device using single plane port radiographs and a remote computed tomography scanner. Vet Radiol Ultrasound. 2003 Jul-Aug;44(4):470-5. 40. Saw CB, Yakoob R, Enke CA, Lau TP, Ayyangar KM. Immobilization devices for intensity-modulated radiation therapy (IMRT). Med Dosim. 2001 Spring;26(1):71-7. 41. Stock M, Pasler M, Birkfellner W, Homolka P, Poetter R, Georg D. Image quality and stability of image-guided radiotherapy (IGRT) devices: A comparative study. Radiother Oncol. 2009 Oct;93(1):1-7. 42. Webster GJ, Rowbottom CG, Mackay RI. Accuracy and precision of an IGRT solution. Med Dosim. 2009 Summer;34(2):99-106. 43. Ruchala KJ, Olivera GH, Schloesser EA, Mackie TR. Megavoltage CT on a tomotherapy system. Phys Med Biol. 1999 Oct;44(10):2597-621. 44. Letourneau D, Wong JW, Oldham M, Gulam M, Watt L, Jaffray DA, Siewerdsen JH, Martinez AA. Cone-beam-CT guided radiation therapy: technical implementation. Radiother Oncol. 2005 Jun;75(3):279-86. 45. Xing L, Thorndyke B, Schreibmann E, Yang Y, Li TF, Kim GY, Luxton G, Koong A. Overview of image-guided radiation therapy. Med Dosim. 2006 - 109 - Summer;31(2):91-112. 46. Purdy JA. 3D treatment planning and intensity-modulated radiation therapy. Oncology (Williston Park). 1999 Oct;13(10 Suppl 5):155-68. 47. Aird EG, Conway J. CT simulation for radiotherapy treatment planning. Br J Radiol. 2002 Dec;75(900):937-49. 48. Wood KA, Hoskin PJ, Saunders MI. Positron emission tomography in oncology: a review. Clin Oncol (R Coll Radiol). 2007 May;19(4):237-55. 49. van Baardwijk A, Baumert BG, Bosmans G, van Kroonenburgh M, Stroobants S, Gregoire V, Lambin P, De Ruysscher D. The current status of FDG-PET in tumour volume definition in radiotherapy treatment planning. Cancer Treat Rev. 2006 Jun;32(4):245-60. 50. Fox JL, Rengan R, O'Meara W, Yorke E, Erdi Y, Nehmeh S, Leibel SA, Rosenzweig KE. Does registration of PET and planning CT images decrease interobserver and intraobserver variation in delineating tumor volumes for non-small-cell lung cancer? Int J Radiat Oncol Biol Phys. 2005 May 1;62(1):70-5. 51. Breen SL, Publicover J, De Silva S, Pond G, Brock K, O'Sullivan B, Cummings B, Dawson L, Keller A, Kim J, Ringash J, Yu E, Hendler A, Waldron J. Intraobserver and interobserver variability in GTV delineation on FDG-PET-CT images of head and neck cancers. Int J Radiat Oncol Biol Phys. 2007 Jul 1;68(3):763-70. 52. Townsend DW, Beyer T, Blodgett TM. PET/CT scanners: a hardware approach to image fusion. Semin Nucl Med. 2003 Jul;33(3):193-204. - 110 - 53. Brambilla M, Matheoud R, Secco C, Loi G, Krengli M, Inglese E. Threshold segmentation for PET target volume delineation in radiation treatment planning: the role of target-to-background ratio and target size. Med Phys. 2008 Apr;35(4):1207-13. 54. Juweid ME, Cheson BD. Positron-Emission Tomography and Assessment of Cancer Therapy. N Engl J Med. 2006 February 2, 2006;354(5):496-507. 55. Kessler ML. Image registration and data fusion in radiation therapy. Br J Radiol. 2006 Sep;79 Spec No 1:S99-108. 56. Schoder H, Yeung HW, Gonen M, Kraus D, Larson SM. Head and neck cancer: clinical usefulness and accuracy of PET/CT image fusion. Radiology. 2004 Apr;231(1):65-72. 57. Deniaud-Alexandre E, Touboul E, Lerouge D, Grahek D, Foulquier JN, Petegnief Y, Gres B, El Balaa H, Keraudy K, Kerrou K, Montravers F, Milleron B, Lebeau B, Talbot JN. Impact of computed tomography and 18F-deoxyglucose coincidence detection emission tomography image fusion for optimization of conformal radiotherapy in non-small-cell lung cancer. Int J Radiat Oncol Biol Phys. 2005 Dec 1;63(5):1432-41. 58. Pluim JP, Maintz JB, Viergever MA. Mutual-information-based registration of medical images: a survey. IEEE Trans Med Imaging. 2003 Aug;22(8):986-1004. 59. Erdi YE, Wessels BW, DeJager R, Erdi AK, Der L, Cheek Y, Shiri R, Yorke E, Altemus R, Varma V, et al. A new fiducial alignment system to overlay abdominal - 111 - computed tomography or magnetic resonance anatomical images with radiolabeled antibody single-photon emission computed tomographic scans. Cancer. 1994 Feb 1;73(3 Suppl):923-31. 60. Hill DL, Batchelor PG, Holden M, Hawkes DJ. Medical image registration. Phys Med Biol. 2001 Mar;46(3):R1-45. 61. Cha J, Cofer RH, Kozaitis SP. Extended Hough transform for linear feature detection. Pattern Recognition. 2006;39(6):1034-43. 62. Bar-Shalom R, Yefremov N, Guralnik L, Gaitini D, Frenkel A, Kuten A, Altman H, Keidar Z, Israel O. Clinical performance of PET/CT in evaluation of cancer: additional value for diagnostic imaging and patient management. J Nucl Med. 2003 Aug;44(8):1200-9. 63. Ford EC, Herman J, Yorke E, Wahl RL. 18F-FDG PET/CT for image-guided and intensity-modulated radiotherapy. J Nucl Med. 2009 Oct;50(10):1655-65. 64. Brock KK. Results of a multi-institution deformable registration accuracy study (MIDRAS). Int J Radiat Oncol Biol Phys. 2010 Feb 1;76(2):583-96. 65. Hwang AB, Bacharach SL, Yom SS, Weinberg VK, Quivey JM, Franc BL, Xia P. Can positron emission tomography (PET) or PET/Computed Tomography (CT) acquired in a nontreatment position be accurately registered to a head-and-neck radiotherapy planning CT? Int J Radiat Oncol Biol Phys. 2009 Feb 1;73(2):578-84. 66. Wood KA, Hoskin PJ, Saunders MI. Positron emission tomography in oncology: a - 112 - review. Clin Oncol. 2007 May;19(4):237-55. 67. Juweid ME, Cheson BD. Positron-Emission Tomography and Assessment of Cancer Therapy. N Engl J Med. 2006 February 2, 2006;354(5):496-507. 68. van Baardwijk A, Baumert BG, Bosmans G, van Kroonenburgh M, Stroobants S, Gregoire V, Lambin P, De Ruysscher D. The current status of FDG-PET in tumour volume definition in radiotherapy treatment planning. Cancer Treat Rev. 2006 Jun;32(4):245-60. 69. Bradley J, Thorstad WL, Mutic S, Miller TR, Dehdashti F, Siegel BA, Bosch W, Bertrand RJ. Impact of FDG-PET on radiation therapy volume delineation in non-small-cell lung cancer. Int J Radiat Oncol, Biol, Phys. 2004 May 1;59(1):78-86. 70. Ashamalla H, Guirgius A, Bieniek E, Rafla S, Evola A, Goswami G, Oldroyd R, Mokhtar B, Parikh K. The impact of positron emission tomography/computed tomography in edge delineation of gross tumor volume for head and neck cancers. Int J Radiat Oncol, Biol, Phys. 2007 Jun 1;68(2):388-95. 71. Simon E, Fox TH, Lee D, Waller AF, Pantalone P, Jani AB. PET lesion segmentation using automated iso-intensity contouring in head and neck cancer. Technol Cancer Res Treat. 2009 Aug;8(4):249-55. 72. Gross MW, Weber WA, Feldmann HJ, Bartenstein P, Schwaiger M, Molls M. The value of F-18-fluorodeoxyglucose PET for the 3-D radiation treatment planning of malignant gliomas. Int J Radiat Oncol, Biol, Phys. 1998 Jul 15;41(5):989-95. - 113 - 73. Lin LL, Yang Z, Mutic S, Miller TR, Grigsby PW. FDG-PET imaging for the assessment of physiologic volume response during radiotherapy in cervix cancer. Int J Radiat Oncol, Biol, Phys. 2006 May 1;65(1):177-81. 74. Mah K, Caldwell CB, Ung YC, Danjoux CE, Balogh JM, Ganguli SN, Ehrlich LE, Tirona R. The impact of (18)FDG-PET on target and critical organs in CT-based treatment planning of patients with poorly defined non-small-cell lung carcinoma: a prospective study. Int J Radiat Oncol, Biol, Phys. 2002 Feb 1;52(2):339-50. 75. Horan G, Roques TW, Curtin J, Barrett A. "Two are better than one": a pilot study of how radiologist and oncologists can collaborate in target volume definition. Cancer Imaging. 2006;6:16-9. 76. Van de Steene J, Linthout N, de Mey J, Vinh-Hung V, Claassens C, Noppen M, Bel A, Storme G. Definition of gross tumor volume in lung cancer: inter-observer variability. Radiother Oncol. 2002 Jan;62(1):37-49. 77. Caldwell CB, Mah K, Skinner M, Danjoux CE. Can PET provide the 3D extent of tumor motion for individualized internal target volumes? A phantom study of the limitations of CT and the promise of PET. Int J Radiat Oncol, Biol, Phys. 2003 Apr 1;55(5):1381-93. 78. Li H, Thorstad WL, Biehl KJ, Laforest R, Su Y, Shoghi KI, Donnelly ED, Low DA, Lu W. A novel PET tumor delineation method based on adaptive region-growing and dual-front active contours. Med Phys. 2008 Aug;35(8):3711-21. - 114 - 79. Montgomery DW, Amira A, Zaidi H. Fully automated segmentation of oncological PET volumes using a combined multiscale and statistical model. Med Phys. 2007 Feb;34(2):722-36. 80. Geets X, Lee JA, Bol A, Lonneux M, Gregoire V. A gradient-based method for segmenting FDG-PET images: methodology and validation. Eur J Nucl Med Mol Imaging. 2007 Sep;34(9):1427-38. 81. van Baardwijk A, Bosmans G, Boersma L, Buijsen J, Wanders S, Hochstenbag M, van Suylen RJ, Dekker A, Dehing-Oberije C, Houben R, Bentzen SM, van Kroonenburgh M, Lambin P, De Ruysscher D. PET-CT-based auto-contouring in non-small-cell lung cancer correlates with pathology and reduces interobserver variability in the delineation of the primary tumor and involved nodal volumes. Int J Radiat Oncol Biol Phys. 2007 Jul 1;68(3):771-8. 82. Schinagl DA, Vogel WV, Hoffmann AL, van Dalen JA, Oyen WJ, Kaanders JH. Comparison of five segmentation tools for 18F-fluoro-deoxy-glucose-positron emission tomography-based target volume definition in head and neck cancer. Int J Radiat Oncol Biol Phys. 2007 Nov 15;69(4):1282-9. 83. Aristophanous M, Penney BC, Pelizzari CA. The development and testing of a digital PET phantom for the evaluation of tumor volume segmentation techniques. Med Phys. 2008 Jul;35(7):3331-42. 84. Erdi YE, Mawlawi O, Larson SM, Imbriaco M, Yeung H, Finn R, Humm JL. Segmentation of lung lesion volume by adaptive positron emission tomography - 115 - image thresholding. Cancer. 1997 Dec 15;80(12 Suppl):2505-9. 85. Daisne JF, Sibomana M, Bol A, Doumont T, Lonneux M, Gregoire V. Tri-dimensional automatic segmentation of PET volumes based on measured source-to-background ratios: influence of reconstruction algorithms. Radiother Oncol. 2003 Dec;69(3):247-50. 86. Jentzen W, Freudenberg L, Eising EG, Heinze M, Brandau W, Bockisch A. Segmentation of PET volumes by iterative image thresholding. J Nucl Med. 2007 Jan;48(1):108-14. 87. Drever L, Robinson DM, McEwan A, Roa W. A local contrast based approach to threshold segmentation for PET target volume delineation. Med Phys. 2006 Jun;33(6):1583-94. 88. Drever L, Roa W, McEwan A, Robinson D. Iterative threshold segmentation for PET target volume delineation. Med Phys. 2007 Apr;34(4):1253-65. 89. Black QC, Grills IS, Kestin LL, Wong CY, Wong JW, Martinez AA, Yan D. Defining a radiotherapy target with positron emission tomography. Int J Radiat Oncol, Biol, Phys. 2004 Nov 15;60(4):1272-82. 90. Schaefer A, Kremp S, Hellwig D, Rube C, Kirsch CM, Nestle U. A contrast-oriented algorithm for FDG-PET-based delineation of tumour volumes for the radiotherapy of lung cancer: derivation from phantom measurements and validation in patient data. Eur J Nucl Med Mol Imaging. 2008 Nov;35(11):1989-99. - 116 - 91. Surti S, Karp JS. Imaging characteristics of a 3-dimensional GSO whole-body PET camera. J Nucl Med. 2004 Jun;45(6):1040-9. 92. Du Y, Tsui BM, Frey EC. Partial volume effect compensation for quantitative brain SPECT imaging. IEEE Trans Med Imaging. 2005 Aug;24(8):969-76. 93. Sethian JA. Level Set Methods and Fast Marching Methods. Second ed. New York: Cambridge University Press; 1999. 94. Westerterp M, Pruim J, Oyen W, Hoekstra O, Paans A, Visser E, van Lanschot J, Sloof G, Boellaard R. Quantification of FDG PET studies using standardised uptake values in multi-centre trials: effects of image reconstruction, resolution and ROI definition parameters. Eur J Nucl Med Mol Imaging. 2007 Mar;34(3):392-404. 95. Pal NR, Pal SK. A review on image segmentation techniques. Pattern Recognition. 1993;26(9):1277-94. 96. Yang J, Duncan JS. 3D image segmentation of deformable objects with joint shape-intensity prior models using level sets. Med Image Anal. 2004 Sep;8(3):285-94. 97. Paulino AC, Koshy M, Howell R, Schuster D, Davis LW. Comparison of CT- and FDG-PET-defined gross tumor volume in intensity-modulated radiotherapy for head-and-neck cancer. Int J Radiat Oncol, Biol, Phys. 2005 Apr 1;61(5):1385-92. 98. Crum WR, Hartkens T, Hill DL. Non-rigid image registration: theory and practice. Br J Radiol. 2004;77 Spec No 2:S140-53. - 117 - 99. Ciernik IF, Dizendorf E, Baumert BG, Reiner B, Burger C, Davis JB, Lutolf UM, Steinert HC, Von Schulthess GK. Radiation treatment planning with an integrated positron emission and computer tomography (PET/CT): a feasibility study. Int J Radiat Oncol Biol Phys. 2003 Nov 1;57(3):853-63. 100. Paulino AC, Koshy M, Howell R, Schuster D, Davis LW. Comparison of CT- and FDG-PET-defined gross tumor volume in intensity-modulated radiotherapy for head-and-neck cancer. Int J Radiat Oncol Biol Phys. 2005 Apr 1;61(5):1385-92. 101. Wang D, Schultz CJ, Jursinic PA, Bialkowski M, Zhu XR, Brown WD, Rand SD, Michel MA, Campbell BH, Wong S, Li XA, Wilson JF. Initial experience of FDG-PET/CT guided IMRT of head-and-neck carcinoma. Int J Radiat Oncol Biol Phys. 2006 May 1;65(1):143-51. 102. Cannon DM, Lee NY. Recurrence in region of spared parotid gland after definitive intensity-modulated radiotherapy for head and neck cancer. Int J Radiat Oncol Biol Phys. 2008 Mar 1;70(3):660-5. 103. Nishioka T, Shiga T, Shirato H, Tsukamoto E, Tsuchiya K, Kato T, Ohmori K, Yamazaki A, Aoyama H, Hashimoto S, Chang TC, Miyasaka K. Image fusion between 18FDG-PET and MRI/CT for radiotherapy planning of oropharyngeal and nasopharyngeal carcinomas. Int J Radiat Oncol Biol Phys. 2002 Jul 15;53(4):1051-7. 104. Schwartz DL, Ford EC, Rajendran J, Yueh B, Coltrera MD, Virgin J, Anzai Y, Haynor D, Lewellen B, Mattes D, Kinahan P, Meyer J, Phillips M, Leblanc M, - 118 - Krohn K, Eary J, Laramore GE. FDG-PET/CT-guided intensity modulated head and neck radiotherapy: a pilot investigation. Head Neck. 2005 Jun;27(6):478-87. 105. Adams S, Baum RP, Stuckensen T, Bitter K, Hor G. Prospective comparison of 18F-FDG PET with conventional imaging modalities (CT, MRI, US) in lymph node staging of head and neck cancer. Eur J Nucl Med. 1998 Sep;25(9):1255-60. - 119 -