Respiratory Motion Tracking in Magnetic Resonance Imaging
with Pilot Tone Technology
Thesis
Presented in Partial Fulfillment of the Requirements for the Degree Master of Science in
the Graduate School of The Ohio State University
By
Mary Lenk
Graduate Program in Electrical and Computer Engineering
The Ohio State University
2018
Thesis Committee
Dr. Lee Potter, Advisor
Dr. Rizwan Ahmad
1
Copyrighted by
Mary Lenk
2018
2
Abstract
This thesis explores the hypothesis that Pilot Tone (PT) technology can encode respiratory induced motion of the heart to improve cardiac magnetic resonance (MR) imaging. Pilot tone technology is advantageous due to its high sampling rate to provide high temporal resolution in tracking and predicting respiration. Also, the PT signal has the potential to provide motion information without interrupting the pulse sequence to perform motion compensated scans. A prediction model is hypothesized to account for in-plane and through-plane motions due to respiration.
A proof-of-concept experiment was designed to explore the ability of the PT signal to encode respiratory-induced motion. The PT signal was processed retrospectively offline and compared to a reference for respiratory motion. The two signals had a high correlation and show preliminary success for the PT to detect respiration. A linear filter was then designed to predict motion from a training phase using the same reference signal. The linear filter was successful with peak/trough locations between the prediction and the reference signal having a correlation coefficient of 0.9999 for end-expiration and end-inspiration prediction.
Furthermore, a PT transmitter was designed and constructed for implementation of additional experiments. The transmitter was designed to be programmable, battery-
ii powered, MR-safe, and portable for placement at various locations in the bore during scans.
iii
Dedication
To my family and friends for their love and support.
iv
Acknowledgments
In working on this thesis I have had the pleasure of working with many brilliant individuals. I would first like to thank my advisor Dr. Lee Potter for encouraging me to pursue this research and sharing his knowledge of signal processing throughout my undergraduate and graduate studies. This thesis came about as an extension to the work done by Michael Bush, and I’d like to thank him for providing me with the appropriate background knowledge and goals to drive the project. I would also like to thank Dr.
Rizwan Ahmad and the CMR research group at The Ohio State University (OSU) for providing their expertise in cardiac MR and guidance on the project. Also, Yingmin Liu was very helpful in conducting scans for the experiments. In designing the transmitter, the antenna design was created by Xiozhen Yang and Dr. Villoroel in the Electrical and
Computer Engineering department at OSU. The transmitter was developed with the help of Siddarth Baskar who further designed the electronics layout for the printed circuit board. Finally, I’d like to thank my family and friends for their unconditional support.
This work was supported by the National Science Foundation under grant IIP-1539961.
v
Vita
2016 ...... B.S. Electrical and Computer Engineering,
The Ohio State University
2017 to 2018 ...... Graduate Research Associate,
The Ohio State University
Fields of Study
Major Field: Electrical and Computer Engineering
vi
Table of Contents
Abstract ...... ii Dedication ...... iv Acknowledgments ...... v Vita ...... vi Table of Contents ...... vii List of Tables ...... viii List of Figures ...... ix Chapter 1. Introduction ...... 1 Chapter 2. Signal Model ...... 7 Chapter 3. Signal Processing ...... 10 3.1 Preprocessing ...... 11 3.2 Linear Predictive Coding ...... 13 Chapter 4. Transmitter Design ...... 16 Chapter 5. Experimentation ...... 20 5.1 Experiment 1, Proof-of-Concept ...... 20 5.2 Experiment 2, Transmitter Performance...... 29 Chapter 6. Discussion ...... 34 Chapter 7. Future Work and Conclusion ...... 35 References ...... 37 Appendix A. Transmitter Antenna ...... 40 Appendix B. Transmitter Electronics ...... 44 Appendix C. Matlab Code ...... 47 Appendix D. Programming Transmitter ...... 60
vii
List of Tables
Table 1: Antenna Dimensions ...... 40 Table 2: Transmitter bill of materials...... 46
viii
List of Figures
Figure 1: Description of PROCO (10)...... 3 Figure 2: High-level description of PROMPT (10)...... 5 Figure 3: Data acquisition for the training phase (10)...... 5 Figure 4: Example of PT in image domain (10)...... 8 Figure 5: Block diagram of PT path for respiratory training...... 10 Figure 6: Block diagram of system H...... 12 Figure 7: Derived Weiner-Hopf filter coefficients for a sinusoid with AWGN...... 15 Figure 8: Block diagram of pilot tone transmitter...... 18 Figure 9: Pilot tone transmitter with monopole antenna...... 18 Figure 10: Reverse side of pilot tone transmitter with lithium-ion battery...... 19 Figure 11: Experimental setup for proof-of-concept experiment (10)...... 21 Figure 12: FFT of raw image data in read-out direction ...... 22 Figure 13: Absolute value of raw pilot tone data for 9 receive channels...... 22 Figure 14: Low pass filter of pilot tone with cutoff at 0.75 Hz...... 23 Figure 15: MOCO region of interest highlighted at ribcage...... 24 Figure 16: SVD of filtered pilot tone compared to MOCO signal...... 25 Figure 17: Comparing peak and trough location for pilot tone and MOCO signal...... 26 Figure 18: Stem plot of L = 100 trained filter coefficients for each ICA vector...... 27 Figure 19: Prediction results from applying filter coefficients to previous 100 samples. 28 Figure 20: Peak and trough locations for predictions results...... 29 Figure 21: Programming of pilot tone transmitter using Eval Kit...... 30 Figure 22: Imaging of phantom with pilot tone...... 31 Figure 23: Primary singular vector of demodulated PT signal...... 33 Figure 24: Meander line antenna design...... 40 Figure 25: Antenna simulation circuit model...... 41 Figure 26: Antenna simulation S11 results...... 42 Figure 27: Simulation gain pattern...... 42 Figure 28: Simulation realized gain...... 43 Figure 29: 3-D realized gain...... 43 Figure 30: Transmitter electronics schematic (Part 1)...... 44 Figure 31: Transmitter electronics schematic (Part 2)...... 45 Figure 32: Eval Kit GUI prompt screen...... 60 Figure 33: Eval Kit Software GUI ...... 61
ix
Chapter 1. Introduction
Cardiac Magnetic Resonance (CMR) is an imaging technique that can provide a comprehensive evaluation of the cardiovascular system. Imaging of the cardiovascular system requires compensation of motion from both the respiratory and the cardiac systems to provide diagnostic images. Gating techniques are often used to acquire images restricted to a certain phase window for respiratory and cardiac motion to limit the effects caused by motion. These techniques require tracking the motion from each source.
Respiration can be compensated using several techniques. The simplest way to eliminate breathing artifacts is to ask patients to hold their breath. However, using breath- holds is challenging in a clinical environment since many patients cannot hold their breath for the 20-40 seconds needed to acquire diagnostic quality images. Using breath- holds also limits the types of scans that can be performed. Another option to track respiration is through the use of respiratory bellows. While respiratory bellows are an attractive option for creating a motion-model, they do not directly observe the heart motion variation in breathing patterns that could create inaccuracies in the model. Also, additional equipment and set up time is required to use respiratory bellows, which increases discomfort to the patients and requires more scanner time. Motion has also been tracked using self-navigation from oversampling of k-space and retrospectively binning
1 data into motion states or through coil clustering (1,2). An additional technique explores measuring thermal noise covariances among receivers (3).
A popular technique for handling respiration is through navigator (NAV) echoes.
NAV echoes use additional radio frequency (RF) pulses in the acquisition sequence to track the liver dome (4). Navigator echoes can be used either retrospectively or prospectively to gate images to acquisition periods during end-expiration in the respiratory cycle where breathing-induced motion of the heart is reduced (5,6). For single-shot applications, retrospective in-plane motion correction (MOCO) has been used instead of navigators to eliminate respiratory motion from images. MOCO, however, does not correct for through-plane motion. Prospective motion compensation (PROCO) is an attractive alternative to MOCO because it can account for through plane motion.
Figure 1 depicts the concept of PROCO. In traditional free-breathing acquisitions without PROCO (first row), the respiratory induced motion of the heart changes the content of the imaging plane (dotted line). In contrast, PROCO (second row) tracks the imaging plane to counteract the breathing motion, capturing the same anatomical location across different heartbeats.
Current PROCO methods rely on one or more navigator echoes to capture respiratory motion and rely on simple parametric models that are inadequate to describe complex respiratory-induced cardiac motion (RIC) (7,8).
2
Figure 1: Description of PROCO (10).
Navigator echoes face several limitations. The NAV signal treats respiratory motion as a one-dimensional signal which limits accuracy in cases of irregular breathing patterns. Further, cardiac gating restricts acquisition to intervals between R-waves which limits the sampling rate of the NAV to one sample per heartbeat. This sampling of the
NAV pulses interrupts the pulse sequence block which further reduces acquisition efficiency. Utilizing NAV pulses also limits flexibility of scans because it is incompatible or inefficient for many CMR protocols. For example, in cine imaging, RF pulses are continuously being acquired over the entire RR interval and cannot afford 70 ms to place a navigator echo within the sequence. In addition, the NAV requires a training period to monitor end-expiration position which incurs additional scan time.
A viable solution to compensate for the drawbacks of the NAV would be to use a signal that would not interrupt the acquisition pulse sequence. The goal of this research is to explore the potential of a new framework to prospectively compensate respiratory motion. The proposed method, called PROspective Motion compensation using Pilot 3
Tone (PROMPT) (10) employs Pilot Tone (PT) technology (9) which provides a reliable surrogate measurement of respiratory motion (10).
Pilot Tone technology uses an external RF signal transmitted close to the central frequency of the MRI scanner. It is hypothesized that producing a small amplitude PT with a fixed frequency within the receive frequency band can be used to monitor respiratory motion (9). Further work with PT technology has explored its ability to track respiration in 2-dimensions using PT and to track cardiac motion (11-13).
This thesis explores the ability of pilot tone to track and model respiration for the
purpose of prospective motion compensation. Signal processing and an adaptive model
will be developed to predict the breathing patterns to assist in creating motion
compensated images and replace the NAV signal for higher resolution, dimensionality, and flexibility in scans. During a training phase, a series of low-resolution images will be collected across several heart beats while pilot tone is continuously sampled. The motion
parameters of the heart across the collected images will be used as training data for the
tracker. Once a model is trained, the PT signal can predict respiratory motion, from
which gradient adjustment can be performed prospectively during data acquisition.
Descriptions of the training and application phase can be seen below in Figures 2 and 3.
4
Figure 2: High-level description of PROMPT (10).
Figure 3: Data acquisition for the training phase (10).
During the training phase pictured in Figure 3, three single-shot, low resolution, orthogonal views (V1, V2, V3) will be collected from each heartbeat. The respiratory 5 motion will be captured by the multichannel PT signal, which is continuously collected.
For simplicity, the PT signal from three channels is depicted. The amplitude and phase variations in the PT signal are hypothesized to encode the respiratory motion.
In our experimentation, we focused on a proof-of-concept study to validate that pilot tone can be used for purposes of prospective motion compensation. Using pilot tone we were able to demonstrate the ability to track respiration with high similarity to a reference signal. Peak and trough locations between the demodulated pilot tone and a reference for respiration had a correlation coefficient of 0.998. Further extension was done to create an adaptive filter for prediction of respiratory motion.
Furthermore, an extension of this work with pilot tone signal would be to use the technology for cardiac motion. Prediction of cardiac motion would eliminate the need for an echocardiogram (ECG) signal which is used for triggering scans. Elimination of the
ECG would eliminate hardware and setup time for scans which would reduce costs and increase comfort for the patient. Further, absence of ECG lead wires would reduce image artifacts.
6
Chapter 2. Signal Model
Magnetic resonance imaging (MRI) is a powerful tool able to create diagnostic images of the body without harm due to radiation. MRI utilizes the magnetic spin property of hydrogen to create images. By applying a strong external field, hydrogen protons align and spin along the field direction. The spin frequency, also known as the
Lamor frequency, is 63.87 MHz for a 1.5T magnetic field. Magnetization can be rotated away from this alignment by applying a small bandwidth radio frequency (RF) pulse to excite a slice. Both frequency and phase gradients are then used to temporarily change the resonant frequency and relative phases of protons. A frequency-encoding gradient is used to specify position within a slice by collecting the gradient echoes which vary in frequency depending on type of tissue within a slice. Furthermore, the phase encoding gradient is used to measure relative phases among protons relative to a reference state.
Each pixel will contain frequency and phase information that is collected into a k-space grid. The Fourier transform is then used to decipher between frequencies and encode spatial information. Details of these principles are further discussed in (15).
Using MRI principles, pilot tone technology has the potential to provide a multi- dimensional respiratory motion signal that can be independently acquired simultaneously with image data without imposing any constraints on pulse sequence characteristics. The pilot tone signal is generated by a separate RF source transmitted through an antenna
7 placed in the scanner room. This requires no additional setup time once the transmitter is placed appropriately. The frequency of the PT signal is close to the Lamor frequency and thus is picked up by the receive coil array placed on the patient. The frequency of the pilot tone can be adjusted to be precisely encoded at a fixed spatial location in the readout direction, ideally outside the physical region of interest for the scan. Oversampling is typically done by doubling the sampling rate in the frequency encode direction to double the field of view (FOV), thereby providing plenty of vacant locations (OS) to place the pilot tone. This oversampling does not increase scan time. In the example in Figure 4, the pilot tone signal was placed away from the image and has further vacant locations due to oversampling.
Figure 4: Example of PT in image domain (10).
The pilot tone will appear as either a dot or a “zipper” depending on the type of scan being implemented. For this research, a GRE sequence and TrueFISP sequence were analyzed. The GRE sequence will appear as a zipper of points (a Fourier series) across
8 the pulse sequence. The TrueFISP sequence uses an alternating sign in modulation of the pilot tone signal which results in the pilot tone appearing as a dot in the image domain.
The amplitude and phase of the tone received by the body coil array is hypothesized to be modulated by respiratory motion. Multiple channels can be used depending on number of coils available during scans. This allows for a multidimensional signal to encode motion information. The sampling rate of the signal is one sample per repetition time (TR) interval which lasts a few milliseconds, allowing for very high temporal resolution in estimation of respiration phase. The received signal is then demodulated to baseband by the MRI software before being displayed in the image domain at the scanner workstation.
The signal model of the pilot tone, s, will be as follows for the ith channel received:
� (�) = � (�)��� 2�(� + Δ�)� + � (�) + � (�) where Ai(t) is the amplitude of the received signal, fc is the Lamor frequency, ∆f is the user defined shift in frequency, ϕi(t) is the time varying phase, and Ni(t) is the noise from the channel.
It is conjectured that the amplitude of the signal, Ai(t), contains physical information about the patient’s motion. Irrespective of the physics of the modulation present in the PT signal, any correlation between respiratory motion and the PT waveform can be exploited to track patient motion.
9
Chapter 3. Signal Processing
The raw signal from motion receive coils is processed to extract the pilot tone data and use it for prediction of respiration. The Fourier Transform of raw k-space data in the phase encode direction reveals the position of the pilot tone along the readout direction. The pilot tone can be used for prediction with comparison to a reference signal during a training phase. Additionally, pilot tone may have potential to be used to monitor cardiac motion for purposes of triggering scans. An overview of the signal path can be seen below in Figure 5.
Figure 5: Block diagram of PT path for respiratory training.
10
In the block diagram, the transmitted pilot tone u(t) passes through the noisy channel and is modulated by the patient’s behavior. The received pilot tone s(t) is a multi- dimensional signal that contains a time-series from each of the receive coils. Systems H and G are used to identify and predict respiratory and cardiac motion, respectively. A reference signal r(t) is used for training of the respiratory prediction filter. An extension to cardiac motion is depicted for purposes of echocardiogram (ECG) triggering.
3.1 Preprocessing
The PT signal is extracted at a given frequency location from the Fourier transform in the frequency encode direction of the raw k-space data. This results in a complex-valued time series si(t) with samples every TR seconds. Demodulation is performed by evaluating the absolute value of the complex signal to analyze the amplitude change which is believed to contain the respiratory motion. Further processing ideas are taken in part from the work of Bacher et al. (13) who have found success in singular vector decomposition and independent component analysis as preprocessing steps.
A training phase of M samples is used on the N channels of pilot tone data.
Singular vector decomposition (SVD) is performed on the M-by-N matrix. This provides a spectral decomposition into orthogonal vectors to reduce dimensionality. This aids in noise reduction and reduces the amount of data used in the prediction algorithm. The
SVD results in a projection matrix, U, to project subsequent Nx1 vector of samples to a desired lower dimensional subspace. The subspace spanned by the highest singular values
11 is used, and lower singular values are discarded. In the case of using 18 receive channels, the dimension is reduced to 6 channels through SVD.
Independent component analysis (ICA) is then performed on the six singular vectors. ICA divides the data into statistically independent vectors. Because motion of the heart is believed to be coming from two independent sources, breathing and cardiac motion, ICA is a naturally viable technique to use. The first four independent component vectors are kept. The first two vectors are believed to contain respiratory motion and the third vector may contain cardiac motion. The preprocessing flow diagram can be seen below in Figure 6 which describes system H in Figure 5.
Figure 6: Block diagram of system H.
Prediction of respiratory motion is done by using singular value decomposition
(SVD), independent component analysis (ICA) and an adaptive filter created through linear predictive coding (LPC). LPC aims to drive the error signal e(t) to zero.
12
3.2 Linear Predictive Coding
To create a predictive model for the purposes of prospective imaging, a causal filter was designed using linear predictive coding. The two respiratory time series are used as inputs to the model which aims to predict the next position of the region of interest.
During the training phase, a motion correction (MOCO) signal is used as a reference for tracking and prediction. The MOCO vector is created by isolating a region of interest within an image frame and measuring displacement of the region across consecutive frames. By choosing a region around the rib cage, sternum motion can be tracked which is highly correlated with respiratory motion. For further extension to gradient adjustment for prospective imaging, MOCO can be used to track heart movement in three orthogonal planes (15).
To perform linear predictive coding, the Weiner-Hopf equations were used. The
Weiner-Hopf filter uses the orthogonality principle to minimize mean squared error between the prediction and the reference vector. For illustration, the equations are derived below with prediction of a stable sinusoid with added white Gaussian Noise (AWGN).
The signal model of interest is as follows:
�(�) = �(�) + �(�) where x(t) is meant to represent a simplified version of the received pilot tone containing a sinusoid indicative of respiration d(t) = Acos(2πft) with frequency f and amplitude A.
Here w(t) represents the additive white Gaussian noise (AWGN). Here the sampled version at sampling period T is used to derive a discrete filter.
13
�[�] = �[�] + �[�] = ����(2����) + �[�]
The goal of the Weiner-Hopf equations is to create an estimate of the reference breathing motion d[n] from a memory stack containing the previous L samples of the observation x[n]. The Weiner-Hopf equations are derived a using linear filter that minimizes the mean squared error between the estimate and the reference. The filter coefficients are learned and applied to the memory stack to predict the next value of the reference vector given the noisy observation.