I wish to acknowledge the following people for their role in the completion of this work:
• My supervisors, Prof. Jesus Ruiz Cabello Osuna and Prof. Ignacio Ro- driguez, for introducing me to the field, and their extreme generousity with their time, knowledge and ideas during the work. Thank you for your guid- ance, support, and patience throughout the course of the research and thesis writing.
• My advisor Dr. Arnoldo Santos, for providing me datasets, guiding me with background related information, for sharing his wealth of experience and knowledge to further my research. Thank you for your valuable collaboration and support during the work.
• My academic tutor at University of Autonoma, Madrid, Prof. Javier Diez Guerra, for constant support and help he provided during the course of this project. I appreciate a lot your helps.
• My colleagues in Advanced Imaging Unit, CNIC, Madrid, especially imaging technicians, for many productive knowledge exchanges, for precious idea sharing and for helping me to obtain necessary data for this project. Thank you for helping me during this time.
• My dear parents and my lovely wife, for their devoted love, encouragement and support, and for their collective energy which was a great motivation in my way. Without you, I could not do this work. Thank you, for being there for me, all the time.
i
Magnetic Resonance Imaging (MRI) is considered to be one of the most widely used imaging techniques in cardiovascular applications. Computational models are continually optimizing the images acquisition and processing pipelines. However, previous computational models of cardiac images faced limitations such as high dependence on assumptions, insufficient accuracy and lack of enough experimental data. The main objective of this work was to propose new techniques to improve imaging and data processing.
1. Pulse Sequence Programming for Fast Imaging: MRI is one of the imaging techniques widely used for the acquisition of the phase contrast images required for the analysis of hemodynamic data. More specifically, PC-MRI obtains velocity and direction of flowing blood. A phase contrast pulse sequence based on GE and with velocity compensation has been developed which allows accurate manipulation of blood flow patterns in small arteries of small animals.
2. Relaxation Time Mapping: In MRI, relaxation times represent specific tis- sue properties that can be quantified with the help of specific imaging strategies.
While there are basic software tools for specific pulse sequences, until now there is no universal software program available to automate the mapping of relaxation times from various types of images. To simplify the search space for the optimum fit, using the partial linear relationship between signal intensity and fitting pa- rameters, an alternative computational tool to compute relaxation times has been proposed and validated.
3. Cardiac Segmentation and Quantification: Evaluation of ventricular volumes and function is considered to be one of the crucial tasks in the management of patients with different heart diseases. Cardiac segmentation of right and left ventricles is necessary when it is intended to study these heart chambers or to extract local densitometric information. A reliable and robust image processing tool to extract cardiovascular MRI derived measurements of both ventricles has been developed and validated during this project.
Keywords: Magnetic Resonance Imaging, Pulse Sequence Programming, Relax- ation Time Mapping, Segmentation, Quantification
La resonancia magn´etica (MRI) se considera una de las t´ecnicas de imagen m´as uti- lizadas en aplicaciones cardiovasculares. Los modelos computacionales optimizan constantemente los procesos de adquisici´on y procesamiento de im´agenes. Sin embargo, los modelos computacionales previos de im´agenes card´ıacas enfrentaron limitaciones tales como una alta dependencia de los supuestos, una precisi´on insu- ficiente y la falta de suficientes datos experimentales. El objetivo principal de este trabajo fue proponer nuevas t´ecnicas para mejorar el procesamiento de im´agenes y datos.
1. Programaci´on de secuencias de pulsos para im´agenes r´apidas: la RM es una de las t´ecnicas de imagen ampliamente utilizadas para la adquisici´on de las im´agenes de contraste de fase necesarias para el an´alisis de datos hemodin´amicos.
M´as espec´ıficamente, la t´ecnica de PC-MRI obtiene la velocidad y la direcci´on de la sangre que fluye. Se ha desarrollado una secuencia de pulsos que utiliza im´agenes de contraste de fase utilizando secuencias basadas en GE y la compensaci´on de velocidad permite la manipulaci´on precisa de los patrones de flujo sangu´ıneo en arterias peque˜nas.
2. Mapa de tiempos de relajaci´on: en la RM, los tiempos de relajaci´on representan propiedades caracter´ısticas del tejido que se pueden cuantificar con la ayuda de estrategias de imagen espec´ıficas. Si bien existen herramientas de software b´asicas para secuencias de pulsos espec´ıficas, hasta ahora no hay un pro- grama de software universal disponible para automatizar el mapeo de los tiempos de relajaci´on de varios tipos de im´agenes. Para simplificar el espacio de b´usqueda para el ajuste ´optimo, utilizando la relaci´on lineal parcial entre la intensidad de la se˜nal y los par´ametros de ajuste, se ha propuesto y validado una herramienta computacional alternativa para calcular los tiempos de relajaci´on.
3. Segmentaci´on y cuantificaci´on card´ıacas: la evaluaci´on de los vol´umenes y la funci´on ventricular se considera una de las tareas cruciales en el tratamiento de pacientes con diferentes enfermedades card´ıacas. La segmentaci´on card´ıaca de los ventr´ıculos derecho e izquierdo es necesaria cuando se pretende estudiar estas c´amaras del coraz´on o extraer informaci´on densitom´etrica local. Durante este proyecto se desarroll´o y valid´o una herramienta confiable y robusta basada en el procesamiento de im´agenes para extraer mediciones de ambos ventr´ıculos derivadas de IRM cardiovascular.
Palabras clave: Im´agenes por Resonancia Magn´etica, Programaci´on de Secuen- cias de Pulsos, Mapas de Tiempo de Relajaci´on, Segmentaci´on, Cuantificaci´on
Acknowledgements i
Abstract ii
Resumen iii
List of Figures viii
List of Tables xi
Abbreviations xii
1 Introduction 1
1.1 Imaging . . . 2
1.1.1 Pulse sequence development . . . 2
1.1.2 Magnetic Resonance Imaging Sequences . . . 4
1.1.2.1 Spin Echo Sequences . . . 4
1.1.2.2 Gradient Echo Sequences . . . 6
1.1.3 Considerations and Challenges . . . 7
1.1.3.1 Spatial and Temporal Resolution . . . 7
1.1.3.2 Imaging artifacts . . . 8
1.1.3.3 Dimensions . . . 9
1.1.3.4 Acquisition Time . . . 9
1.2 Analysis (processing and postprocessing tools) . . . 11
1.2.1 Cardiac Parametric Mapping . . . 11
1.2.1.1 Overview of Curve Fitting Methods. . . 11
1.2.1.2 Considerations and Challenges . . . 13
1.2.2 Cardiac Segmentation and Quantification. . . 14
1.2.2.1 Cardiac Imaging Planes . . . 15
1.2.2.2 Overview of Segmentation Methods . . . 17
1.2.2.3 Considerations and Challenges . . . 19
1.3 List of Tools, Equipment and Software . . . 20 iv
1.4 Objectives . . . 21
1.5 Thesis Outline. . . 22
2 Fast, High Spatial and High Resolution Magnetic Resonance Imag- ing of Small Animal Vessels using a Phase Contrast based Pulse Sequence 24 2.1 Introduction . . . 24
2.2 Materials and Methods . . . 27
2.2.1 Animal Model . . . 27
2.2.2 Magnetic Resonance Imaging Protocol . . . 28
2.2.3 Developed Pulse Sequence (Gradient Echo with Multi Slice and Flow Compensation) . . . 31
2.2.3.1 Gradient Echo Module . . . 33
2.2.3.2 Motion Sensitizing Modules . . . 33
2.2.3.3 Phase Difference Reconstruction . . . 38
2.2.3.4 Cardiac Triggering . . . 40
2.2.4 Validation Criteria . . . 42
2.2.4.1 SNR Analysis . . . 42
2.2.4.2 Vessel Segmentation and Registration . . . 43
2.2.4.3 Customizing Sequence Parameters . . . 43
2.2.4.4 Common Measuremens. . . 43
2.3 Results . . . 45
2.3.1 SNR Analysis . . . 45
2.3.2 Image Segmentation and Registration . . . 45
2.3.3 Customizing Sequence Parameters. . . 46
2.3.4 Common Measurements . . . 49
2.4 Discussion . . . 55
3 Linear Curve Fitting of Relaxation Time Maps in Magnetic Res- onance Imaging 58 3.1 Introduction . . . 58
3.2 Materials and Methods . . . 61
3.2.1 Magnetic Resonance Imaging Protocol . . . 61
3.2.2 Proposed Algorithm . . . 64
3.2.2.1 Curve Fitting . . . 65
3.2.2.2 Error Reduction . . . 68
3.2.2.3 Motion Correction (MOLLI T1 images) . . . 68
3.2.3 Validation Criteria . . . 71
3.2.3.1 Comparison with Levenberg-Marquardt and Nelder- Mead . . . 71
3.2.3.2 Linear regression and modified Bland-Altman Anal- ysis . . . 74
3.3 Results . . . 74
3.3.1 Method validation using comparison with Levenberg-Marquardt and Nelder-Mead . . . 74
3.3.2 Method validation using linear regression and modified Bland-
Altman analysis . . . 80
3.4 Discussion . . . 84
4 Automatic 4D Cardiac Segmentation and Quantification to Eval- uate Ventricular Volume and Function in short-axis cine MR Im- ages 86 4.1 Introduction . . . 86
4.2 Materials and Methods . . . 90
4.2.1 Cardiac MR Data . . . 90
4.2.2 Proposed Algorithm . . . 90
4.2.2.1 Preprocessing . . . 91
4.2.2.2 Heart Chamber Delineation . . . 95
4.2.2.3 Image Segmentation . . . 96
4.2.2.4 Removing papillary muscles and trabeculae . . . . 102
4.2.2.5 Extracting LV and RV . . . 102
4.2.2.6 Measurements. . . 103
4.2.3 Validation Criteria . . . 105
4.2.3.1 Ground truth: manual segmentation protocol . . . 105
4.2.3.2 Automatic vs Manual segmentation: Distance Met- rics analysis . . . 107
4.2.3.3 Automatic vs Manual segmentation: Receive Op- erator Characteristic (ROC) analysis . . . 109
4.3 Results . . . 111
4.3.1 Bland-Altman and Linear Regression Ejection Fraction anal- ysis . . . 111
4.3.2 Automatic vs Manual segmentation: Distance Metrics analysis112 4.3.3 Automatic vs Manual segmentation: Receive Operator Char- acteristic (ROC) Analysis . . . 117
4.4 Discussion . . . 118
5 Discussion and Conclusions 120 5.1 Fast, High Spatial and High-Resolution Magnetic Resonance Imag- ing of Small Animal Vessels using a Phase Contrast based Pulse Sequence . . . 120
5.2 Linear Curve Fitting of Relaxation Time Maps in Magnetic Reso- nance Imaging. . . 122 5.3 Automatic 4D Cardiac Segmentation and Quantification to Evalu-
ate Ventricular Volume and Function in short-axis cine MR Images 124
A Appendix 1: pulse sequence user manual and source code 126 B Appendix 2: relaxation time mapping tool source code 160
C Appendix 3: ventricle segmentation tool source code 193
Bibliography 219
1.1 Example of spin echo pulse sequence demonstrating multislice vol-
ume acquisition. . . 5
1.2 Example of gradient echo pulse sequence demonstrating GRE with EPI factor 3. . . 6
1.3 An example of relaxation time mapping using inversion recovery sequence for T1 mapping. . . 12
1.4 An example of curve fitting using linear and non-linear regression. . 13
1.5 Some common cardiac MRI planes. . . 17
1.6 An example of medical image segmentation (using one of the seg- mentation tools developed in this project). . . 18
2.1 Planning of the first 1.1 mm cut for reproducible selection of the pulmonary artery in both rat and mouse. . . 29
2.2 Image planning result showing the section of the main branch of the pulmonary artery and the first branch. . . 30
2.3 Localization of the plane perpendicular to the right branch of the pulmonary artery, immediately after the first bifurcation positioned below the ascending aorta (arrow). . . 30
2.4 Flow maps obtained from the planes selected in Fig. 2.2 and Fig. 2.3. . . 31
2.5 Pulse sequence diagram of Gradient Echo module. . . 34
2.6 Pulse sequence diagram of flow encoding module. . . 37
2.7 Using Gradient Moment nulling module to compensate for flow and motion. . . 38
2.8 Illustration of flow encoding vs flow compensation modules.. . . 39
2.9 Prospective ECG triggering. . . 41
2.10 An example of SNR measurement using magnitude images. . . 46
2.11 Summary of SNR measurement. . . 47
2.12 An example of automatic segmentation of Regions of Interest. . . . 47
2.13 An example of coupling magnitude and phase images and co-registering the ROIs needed for flow analysis. . . 48
2.14 Example of acquiring images for an individual subject with different VENC values. . . 49
2.15 Example of adjustment of the number of the frames in the sequence. 50 2.16 Comparison of flow-time curves extracted for one subject. . . 51
viii
2.17 Comparison of average total flow versus time between hypoxia and
control group in MPA position. . . 51
2.18 Comparison of average total flow versus time between hypoxia and control group in RPA position.. . . 52
2.19 Summary of velocity analysis (min, max and mean) for MPA (top) and RPA (bottom) positions. . . 53
2.20 Plots demonstrating distribution of the flows for hypoxia and con- trol groups in MPA and RPA position. . . 53
2.21 Measurement of flow and cross-sectional area in the main pulmonary artery of a mouse during early systole. . . 55
3.1 Standard Inversion Recovery pulse sequence. . . 63
3.2 Standard Modified Look Locker (MOLLI) sequence. . . 64
3.3 Error reduction of linear curve fitting for T1 MOLLI mapping in small animal study. . . 68
3.4 An example of calculating parametric T1 relaxation time map. . . . 71
3.5 Summary of the proposed method to get parametric relaxation times. 72 3.6 T1 map of a mouse brain (center slice) obtained using the algorithm described in this thesis. Corresponds to test 1. . . 75
3.7 T1 map of a simulated brain (center slice) obtained using the algo- rithm described in this thesis. Corresponds to test 6. . . 76
3.8 Results of quotient index corresponding to tests 1 to 7. . . 77
3.9 Results of A, B and C indices corresponding to tests 1 to 7. . . 77
3.10 Results of Rtavg index corresponding to tests 1 to 7. . . 78
3.11 Two-dimensional plot for test 1. It is apparent that T1 obtained by Bruker’s fit is greater than the one obtained by the fit described in this thesis. . . 79
3.12 Two-dimensional plots for test 2-7. For tests 2-5, It is evident that T1 obtained by the two fitting methods are very similar. For test 6, T1 obtained by Matlab’s fit is lower than the one obtained by the fit described in this thesis. For test 7, T1 obtained by the two fitting methods are similar, even if there is more dispersion than in plots of tests 2–5 and T1 obtained by Matlab’s fit is greater than the one obtained by the fit described in this thesis. . . 80
3.13 Fitting error (y) as a function of T1 (x) after parameters A and B have been fitted. . . 81
3.14 An example image used for method validation using linear regres- sion and modified Bland-Altman analysis. T2 signal of superpara- magnetic Iron Oxide Nanoparticles (SPION ) were evaluated using proposed method. . . 82
3.15 Coronal view for T2-weighted gradient echo 3D MRI of the breast tumor area in a mouse 2h post-injection of IONPs. . . 83
3.16 Bland-Altman analysis results corresponding to T1, T2 and T2* re- laxation time mapping. . . 83
4.1 Representative cardiac cine MR short-axis image . . . 91
4.2 Example of denoising filters applied to the images.. . . 94
4.3 Example of applying selected denoising approach to an image. . . . 95
4.4 Delineation of heart chamber using proposed method for one patient and in a mid-ventricular position. . . 97
4.5 Example of applying segmentation operator to an image. . . 101
4.6 Inclusion (left) and Exclusion of papillary muscles and trabeculae in the analysis. . . 102
4.7 Endocardial contours for one patient . . . 103
4.8 Flow scheme for proposed algorithm. . . 105
4.9 an example of manual delineation of cardiac segmentation . . . 107
4.10 Ejection Fraction Bland-Altman plots for LV. The horizontal lines show average of manual and automatic EF calculations, mean and ±1.96 standard deviation of difference. Vertical axis represents the difference range for two methods. . . 112
4.11 Ejection Fraction Bland-Altman plots for RV. The horizontal lines show average of manual and automatic EF calculations, mean and ±1.96 standard deviation of difference. Vertical axis represents the difference range for two methods. . . 113
4.12 Ejection Fraction Linear Regression plot for LV. . . 114
4.13 Ejection Fraction Linear Regression plot for RV. . . 114
4.14 Distance metrics analysis results summary for basal pixels. . . 115
4.15 Distance metrics analysis results summary for mid ventricular pixels.115 4.16 Distance metrics analysis results summary for apical pixels. . . 116
A.1 source code classes and their relationships. . . 146
A.2 Scan Parameters panel . . . 148
A.3 Flow Compensation panel. . . 149
A.4 Flow Encoding panel. . . 150
A.5 Sequence panel. . . 151
1.1 Common imaging parameters of Spin Echo sequences. . . 5 2.1 Image Acquisition parameters. . . 32 2.2 Flow compensation (FC ) and flow encoding (FE ) scan acquisitions
description. . . 40 2.3 Velocity analysis for MPA position. . . 54 2.4 Velocity analysis for RPA position. . . 54 3.1 Some MRI signal models which can be fitted using the algorithm
described in this chapter. In all cases, the fit can be partially lin- earized so only one parameter which is relaxation time needs to be varied to find the lowest error. . . 67 3.2 List of performed tests. . . 75 4.1 List of available measurements after segmentation.. . . 104 4.2 Criteria for interpreting the results of Automatic vs. Manual seg-
mentation analysis using distance metrics. . . 108 4.3 Calculated indices for ROC analysis. . . 117
xi
1D 1(one) Dimensional 1.5T 1.5(one and half) Tesla 2D 2(two) Dimensional 3D 3(three) Dimensional 3T 3(three) Tesla
4CH 4(four) CHamber 4D 4(four) Dimensional 7T 7(seven) Tesla Acq Acquisition ApoE Apoliprotein E
ASL Arterial Spin Labelling AV Atrio Ventricular
CCM Connected Component Labeling CFD Computational Fluid Dynamics CHD Coronary Heart Disease
CMC Carboxy Methyl Cellulose CMR Cardiac Magnetic Resonance
CNIC Centro Nacional de Investigaciones Cardiovasculares
CO Cardiac Output
cP centi Poise
CSF Cerebro Spinal Fluid
CT Computed Tomography
CV Cardio Vascular
DICOM Digital Imaging and COmmunications in Medicine xii
ECG Electro Cardio Gram ECM Extra Cellular Matrix ECV Extra Cellular Volume
ED End Diastolic
EDV End Diastolic Volume EF Ejection Fraction EPI Echo Planar Imaging ES End Systolic
ESV End Systolic Volume
EX EXcitation
FE Flow Encoding
FE Frequency Encoding FFT Fast Fourier Transform FID Free Induction Delay FOV Field Of View
GE Gradient Echo
GMN Gradient Moment Nulling GRE GRadient Echo
HE Hematoxylin Eosin HLA Horizontal Long Axis
HR Heart Rate
IONP Iron Oxide Nano Particle IR Inversion Recovery
IACUC Institutional Animal Care Use Committee ITN Initial Training Network
KO Knock Out
LDL Low Density Lipoprotein LM Levenberg Marquardt LUT Look Up Table LV Left Ventricle
LVOT Left Ventricular Outflow Tract
MAE Mean Absolute Error MIC Medical Image Computing MM Mathematical Morphology
MOLLI MOdified Look Locker Inversion recovery MRI Magnetic Resonance Imaging
NIH National Institute of Health
NP Nano Particle
PA Pulmonary Artery
PAH Pulmonary Arterial Hypertension PC Phase Contrast
PE Phase Encoding
PET Positron Emission Tomography PRS Phase Read Slice
PSNR Peak Signal to Noise Ratio PWV Pulse Wave Velocity RAM Random Access Memory
RARE RApid imaging with Refocused Echoes
Re Reynolds
RF Radio Frequency
ROI Region Of Interest
RMSE Root Mean Square Error RPS Read Phase Slice
RV Right Ventricle
RVOT Right Ventricular Outflow Tract
SA Short Axis
SCMR Society of Cardiovascular Magnetic Resonance SD Standard Deviation
SE Slice Encoding
SE Spin Echo
SE Structuring Element SF Smoothing Factor
SMC Smooth Muscle Cell SNR Signal to Noise Ratio
SPECT Single Photon Emission Computed Tomography SPR Slice Phase Read
SSE Single Spin Echo
SSFP Steady State Free Precision
SV Stroke Volume
TE Time Echo
TI Time Inversion TR Time Repetition
VEGF Vascular Endothelial Growth Factor VENC Velocity Encoding
VLA Vertical Long Axis
VK Von Kossa
Introduction
Magnetic Resonance Imaging (MRI) is an imaging modality which is used heavily in clinical practice and research to assess different physiological processes and visualize anatomical structures. In the presence of an external magnetic field, this technique uses magnetic properties of the atoms to form the spin concentration and the images. Based on relaxation times of the spins describing the times needed for their recovery after their excitation with a radiofrequency (RF pulse), images with different contrasts can be generated. Due to the non-invasive nature of MRI and its better contrast of soft tissues in comparison with other modalities such as CT, X-ray or PET, it is widely used nowadays for imaging and visualization of different organs and pathological tissues[1].
There are a wide variety of research performed in recent years to improve the quality of the images and the processing based on MR signals. It mainly includes two different categories:
1. Imaging (pulse sequence development): Magnetic resonance images of subjects are acquired by applying a combination of different radiofrequency (RF ) pulses and gradient waveforms inside a magnetic field. Magnetic Res- onance Imaging includes and is not limited to basic anatomical acquisitions and diagnostic techniques such as diffusion, perfusion, or functional imaging.
1
The ongoing research in this area tries to improve the acquisition speed, im- age quality, and contrast. There are currently dozens of available sequences for imaging (T1, T2, T2∗, proton density weighted, diffusion weighting, perfu- sion, magnetization transfer, velocity encoding, etc) and the ongoing research try to optimize them or develop new ones according to specific clinical and research needs[2].
2. Analysis (processing and postprocessing tools): Medical image com- puting (MIC ) is an interdisciplinary field combined of several techniques from fields of computer science, electrical engineering, physics, mathemat- ics, and medicine. MIC focuses on the computational analysis of the medical images and not their acquisition. It tries to apply mathematical and compu- tation models on top of images. These models then help researchers and clin- icians to extract relevant information or knowledge from medical datasets.
The used methods in MIC can be grouped into various categories: image segmentation, image registration, image-based physiological modeling. Re- garding MRI, MIC tries to improve the accuracy of analysis performed by this imaging technique using automatic or semi-automatic developed tools and algorithms. This includes a wide range of tools used in different anal- ysis pipelines: to name some, anatomical studies, quantification of images to extract needed measurements, MRI measurements as complementary to other techniques such as nanoparticles or histological studies, etc[3].
1.1 Imaging
1.1.1 Pulse sequence development
Each MRI scanner has some basic hardware components[4]:
• B0 (main magnetic field): The strength of this field is fixed (1.5, 3, 7T for instance).
• RF field (B1): A time-varying radiofrequency (RF ) field should first be applied to the spin system. This field is near to the Larmor frequency (spe- cific resonant of nuclei). Transmission of electrical currents through RF- transmit coils generates the RF field.
• Gradients: By adding (one or several) additional, weaker magnetic fields, signals coming from different regions of the subject under study can be local- ized spatially. These additional magnetic fields are called gradients. In the standard setting, each device has three gradients, called slice selection, phase encoding and frequency encoding. These three components allow the MRI scanner to produce magnetic fields in x, y, and z-direction; hence localize the obtained signal in 3D space:
1.Slice Selection : By altering the frequency of the excitation pulses to match the frequency at the desired slice position, this gradient allows the scanner to select the particular slice of the region of interest.
2.Frequency Encoding : This gradient, also called the readout gradient, is useful for localizing the signal in a particular slice and one dimension (x or y). Different protons have different precession frequencies and this feature is used to split them in one direction, called frequency or readout direction.
3.Phase Encoding : This gradient’s role is to finalize the process of the protons localization, in the third axis by taking into consideration that pro- tons are accumulating some phase shifts while this gradient is turned on. By using that phase shift information, it is possible to encode the protons in the third dimension.
Pulse sequences are displayed in diagrams reflecting gradients and RF pulses shape and time disposition for data acquisition. The diagrams are superimposed if the parameters need to be shown in a single series or they can be divided into different waveforms. In the standard way usually, these diagrams consist of four waveforms:
one for the RF pulses and three subsequent ones showing x, y and z gradients. As it is evident from the name, pulse sequences demonstrate the timing sequence of
the events of RF and gradient pulses[2]. Example of pulse sequence diagrams can be reviewed through Fig. 1.1 and Fig. 1.2.
The main objective of pulse sequence design is to program pulse and gradient events using software. In that sense, each component of MRI hardware is a mod- ule, and then any modification related to individual modules is considered as an event.
1.1.2 Magnetic Resonance Imaging Sequences
Depending on the needed information, two or three-dimensional sequences can be used for image acquisition. In two dimensional sequences, one or multiple sections (slices) are acquired at a time whereas, in three-dimensional ones, a volumetric region is obtained in each cycle of acquisition.
In general, there are two main categories of sequences in MRI : Spin Echo (SE) and Gradient Echo (GE) based sequences[5]. Almost all sequences in the literature are based on one of these categories with some variations in imaging parameters. These variations cause the sequences to have different spatial and temporal resolutions.
1.1.2.1 Spin Echo Sequences
Spin Echo (SE)[6] sequences are made of an excitation pulse and one or several refocusing pulses of 90◦ or 180◦. Due to this characteristic, sequences under this category are usually used for obtaining T1, T2 or proton density weighted images.
Moreover, RF pulses of SE sequences refocus the off-resonance effects, and there- fore these sequences are safer to use when artifacts related to such effects (e.g., magnetic field inhomogeneity or magnetic susceptibility variations) may interfere the acquisition process. An example of Spin Echo based sequence is shown in Fig.
1.1.
RF
TE/2
90
TE/2
180
Acquire
Slice Selection
TR/2 TR
Phase Encoding
Tp
Frequency Encoding
TS TS/2
Figure 1.1: Example of spin echo pulse sequence demonstrating multislice volume acquisition.
Depending on the tissue types, main magnetic field strength and some other tech- nical considerations, typical values of TR and TE values for each category of weighted images are listed in Table 1.1.
Table 1.1: Common imaging parameters of Spin Echo sequences.
Short TE (≤ 20ms) Long TE (≥ 20ms) Short TR (≤ 700ms) T1-weighted not common
Long TR (≥ 2000ms) Proton Density weighted
T2-weighted
Depending on the number of RF refocusing pulses per repetition time interval, SE sequences can also be categorized as single echo or multiple echoes. In standard single echo SE sequence, RF of each TR consists of 90◦ excitation pulse followed by a 180◦ refocusing pulse. On the other hand, in multiple-echo SE sequences, transverse magnetization which forms RF, repeatedly refocus into subsequent SE s by playing additional RF refocusing pulses and as a result, several echoes are pro- duced. This process is called echo train and is used heavily for T2 measurements.
1.1.2.2 Gradient Echo Sequences
Gradient Echo (GE )[7] sequences do not use 180◦ RF refocusing pulses. Instead, gradient reversal on the frequency encoding axis is used to form the echo. In these sequences, TR is short (2-50 ms) because the flip angle of the excitation pulse is normally much less than 90◦, so the required time for T1 recovery is less. Having shorter TR means that these types of sequences can be used for fast imaging.
Moreover, 3D volume acquisition which needs higher acquisition speed is one of the areas in which GE sequences are often used. Some other applications of these sequences are the acquisition of images with bright blood signal, susceptibility weighted images, and applications, which need the breath hold of the patients.
An example of Gradient Echo based sequence is shown in Fig. 1.2.
RF
TE a
TE Acquire
TE
Acquire Acquire
Slice Selection
Phase Encoding
Frequency Encoding
Tp
Figure 1.2: Example of gradient echo pulse sequence demonstrating GRE with EPI factor 3.
By definition, a system is in steady state or dynamic equilibrium, if it is not chang- ing with having constant parameters. In GE sequences, this concept is used fre- quently for both longitudinal and transversal magnetization. Longitudinal mag- netization reaches a steady state after applying a sufficient number of excitation pulses. For transverse magnetization, if response is zero just before each excitation pulse, then the GE sequence is called spoiled and for non-zero values, the sequence is categorized as steady-state free precession (SSFP )[8].
According to the above explanations and several other considerations, there are several variations of GE sequences implementations. Each variation gives quite different contrast in result images. Moreover, spoiled sequences are usually used for 2D interleaved imaging whereas steady state pulse sequences are commonly used for 3D or sequential 2D acquisitions.
1.1.3 Considerations and Challenges
In designing pulse sequences, there are specific considerations for taking into ac- count in order to optimize the use of the sequence. The relevant ones to this work are explained below.
1.1.3.1 Spatial and Temporal Resolution
By definition, Spatial and temporal Resolution refer to precision of imaging with respect to time and geometry characteristics[9].
Spatial resolution generally determines the sharpness of the image and its level of captured details[10]. In MRI, the size of three-dimensional rectangular solids image subunits called voxels defines spatial resolution. Size of these voxels (and therefore resolution) is related to some imaging parameters such as data matrix size, the field of view (FOV ) and slice thickness. Matrix size in typical data acquisition is equal to the number of frequency encoding steps in one direction and number of phase encoding steps in the other direction of the image plane.
Therefore, having all other parameters fixed, increasing the steps in frequency or phase encoding directions gives higher resolution images. It should be however noted that frequency encoding steps depend on how rapidly a Free Induction Decay (FID ) or echo signal is sampled by the scanner and the fact that increasing number of phase encoding steps leads to increase in acquisition period.
Temporal resolution is indicated by the interval between consecutive images. This has especially important significance in cardiac MR (CMR), where the heart is
moving during acquisition, and multiple images during the cardiac cycle should be acquired[10].
1.1.3.2 Imaging artifacts
By definition, MRI artifact is an anomaly seen during the visual representation.
In other words, it is a feature that appears in the MR image and is missing in the original object under study. The sources of these artifacts are different and can be related to subject, imaging hardware or pulse sequence adjustments[11].
Subject related artifacts are the artifacts whose source is the study subject itself.
Motion artifacts are observed because of the motion seen in organs such as the heart or the lungs. Flow artifacts are related to fluid movements such as blood flow. Metal artifacts are observed at tissue level when local magnetic fields distort the external magnetic field.
Imaging hardware artifacts arise from some failure or misadjustments in the hard- ware of the MRI device. RF quadrature artifacts cause bright spots in the image mainly because of improper detector channel operation. Inhomogeneous external magnetic field causes spatial, intensity (or both) distortions and cause B0 arti- facts. Gradient field (B1 ) artifacts are the ones seen in the images because of using imaging gradients in different axis (slice, phase or frequency). RF noise, zero line, star, bound point, and surface coil artifacts are some of the other common sources of artifacts in this category.
Pulse sequence artifacts are the ones manifested because of data sampling and processing strategies defined in the pulse sequence. Chemical shift artifact in the phase encoding or slice selection directions occurs at the fat/water interface.
Wrap around artifact is the result of mismapping of the anatomy that lies outside of the field of view but within the slice volume. The ringing artifact is caused by under-sampling of high-spatial frequencies at sharp boundaries in the image.
In order to have a more accurate MRI measurement, it is crucial to identify the sources of the artifacts and minimize or eliminate their effects in the pulse sequence development and (or) processing pipeline.
1.1.3.3 Dimensions
Depending on the type of pulse sequence and specific needs of applications, two, three or four-dimensional MR images usually are obtained[2].
In 2D mode, imaging volume is divided into different slices, and each image then gives information about each specific slice of interest. In this type of acquisition, 2D planes are excited. There is also a controllable gap between these 2D slices or continuous (non-gapped) acquisition, and by combining these planes, a volumetric reconstruction model could be generated.
In 3D mode, a volumetric acquisition is performed. The main difference with the 2D mode is that in this type of acquisition, the whole slab (volume) is excited and encoded instead of single slices. 3D imaging’s acquisition period is longer and generally has a higher SNR because all the spins inside the slab are contributing to the signal. It is also possible to obtain thinner slices out of 3D data.
In 4D mode, apart from three standard dimensions (x, y, and z), a fourth di- mension (usually time, cardiac or breath cycle) is defined. Then volumetric data are obtained through the fourth dimension, as well. For instance, using 4D flow MRI, information regarding the temporal and spatial evolution of 3D blood flow with full volumetric coverage of any cardiac or vascular region of interest can be generated.
1.1.3.4 Acquisition Time
According to the definition, acquisition time is the amount of needed time to collect all of the data for a particular sequence. This time does not include the needed time to reconstruct the image[2].
In 2D imaging, it is calculated by the formula:
TAcq = NEX × NP E× T R (1.1)
In which
NEX: number of excitations (averages) NP E: phase encoding matrix size T R: repetition time
In 3D imaging, it is calculated by the formula:
TAcq(3D) = NEX × Ny × Nz× T R (1.2)
In which:
NEX: number of excitations (averages) Ny: in-plane phase encoding matrix size Nz: number of slices
T R: repetition time
Therefore, as it is evident from the above formulas, the acquisition time is related to the used pulse sequence and dependent on the assembly of parameters such as repetition time (TR), phase encoding matrix size, number of slices and number of signal averages (NEX).
1.2 Analysis (processing and postprocessing tools)
1.2.1 Cardiac Parametric Mapping
Signal intensity in MRI, depends among other things, on relaxation times. These values are time constants, reflecting the magnetic properties of the regions under assessment. Relaxation times which are routinely measured in spectroscopic or imaging-based MR are: longitudinal or spin-lattice relaxation time (T1), transver- sal or spin-spin relaxation time (T2) and apparent transversal relaxation time (T2∗)[12][13][14]. The type of pulse sequence and also the technical parameters used during the acquisition determine the degree that each time factor affects the final signal intensity. For a detailed description of MRI signal formation, the reader is referred to the textbooks on this topic[4][2][15]. Images which represent the values of relaxation times for each pixel are called relaxation time maps. These maps have several applications in research and clinical fields such as myocardial extracellular volume (ECV ) quantification in cardiac MRI [16], oedema detection in acute myocardial infarction[17] and determining early tumour progression in brain[18]. An example of MRI parametric mapping is shown in Fig. 1.3.
1.2.1.1 Overview of Curve Fitting Methods
Regression analysis is the generic name assigned to a group of methods trying to model the relationship between variables. There is usually two type of variables:
response and predictor. It is desirable to see how changes in the response are corresponding to the changes in the predictor variables. It is also possible to use regression models to make predictions based on the values of the predictors.
There are a wide variety of regression methods and based on the type of response variable, desired fitness, data, and estimation method, it is possible to choose between them[19].
Figure 1.3: An example of relaxation time mapping using inversion recovery sequence for T1 mapping.
Reprinted from: A. J. Taylor, M. Salerno, R. Dharmakumar, and M. Jerosch- Herold. T1 mapping: Basic techniques and clinical applications. JACC Cardio-
vasc. Imaging, 9:67-81, 2016.
Curve fitting is the process in which regression analysis is used to find a fit to the specific curves in the dataset (see Fig. 1.4). As a general categorization, there are two groups of techniques in curve fitting: linear and non-linear regression[20].
In linear regression[21], the basic assumption is that the relationship between the dependent variable and predictors are linear. The relationships between variables are modeled using functions called linear predictors. In this model, unknown model parameters are estimated from the data. Linear regression models are often fitted using an approach called least squares. In this method, the best fit minimizes the sum of squared residuals (the difference seen between observed fitted value).
Ridge regression, lasso, bayesian and least angle regression are some other common approaches in this category.
Non-linear regression[22] is a very powerful alternative to linear regression. The main reason behind this is that in non-linear regression, we are not limited to the linear relationship of variables assumption. Instead, several non-linear functions such as exponential, logarithmic, trigonometric, power, Gaussian functions, and Lorenz curves can be examined on top of the dataset in order to find the opti- mal fit for the curve. In order to reach to best fitting accuracy however, it is essential to feed non-linear models with the proper bounds and starting values.
Many implementations of curve fitting using non-linear regression are available in literature[23]. The most common one is non-linear least squares method[24]. In this approach, a set of m observations that is non-linear in response to n unknown parameters (m greater or equal to n) is fitted.
Figure 1.4: An example of curve fitting using linear (left) and non-linear (right) regression.
1.2.1.2 Considerations and Challenges
Although there are several pulse sequences and curve fitting methods already available to use, still there are several limitations and technical considerations to use such options[25].
For T1 mapping for instance, in cardiac studies, the impact of heart rate variation on stress T1 mapping is undeniable. Moreover, factors such as T1 sensitivities to T2, magnetization transfer (MT ) effects and breath-hold duration and motion during acquisition, affect the signal accuracy and therefore T1-maps.
Regarding T2 and T2∗, when using a single spin echo acquisition, the selected TE values should be close enough to T2 of the under-study tissue because if TE is much longer than T2, the fit is incorrectly weighted by the relaxed part of signal inten- sity versus TE. Thermal effects seen in multi-spin echo sequences contribute to dephasing of the spins and make T2 even shorter. With single spin-echo sequences (SSE ), and repeating the acquisition for different echoes, this problem is addressed to a degree. However, this method needs a longer acquisition time. Increasing TR also increases the signal-to-noise ratio (SNR) in relaxometry evaluation.
Regarding curve fitting, since T1 and T2(∗) has non-linear behavior, usually non- linear regression models are used to find the fit for them. The main limitation of using such techniques for finding the best fits is that the domain of possible func- tions to search is wide and so the processing time can be extended. Moreover, the initial setup of such methods (boundaries, assumptions) is critical in the workflow of the algorithms. Finally, in finding the best fits, it is possible that the algorithm gets stuck in local optima and do not give the global one for the fit.
1.2.2 Cardiac Segmentation and Quantification
MRI is considered to be an accurate non-invasive imaging modality for obtaining precise information regarding morphology, perfusion, blood flow, and tissues via- bility. Thanks to recent developments in acquiring high quality and high temporal resolution images using CMR imaging techniques, analyzing cardiac anatomy and physiology now becomes more accurate and reliable. Nowadays, volume quan- tification, ejection and regurgitant fraction analysis and pulmonary arterial flow assessment are between the studies, which can be accomplished using cardiac MRI measurements[26].
With accurately modeling structure and physiology of the left ventricle (LV ), it is possible to understand effects related to heart failure diseases. For instance, intracavitary flow along the time and space dimensions can be extracted from MR images and then be used to analyze blood circulation and/or apex to base me- chanical activation[27]. On the other hand, to study cardiopulmonary disorders such as right ventricular hypertrophy, congenital heart disease, and pulmonary hypertension, occasionally one needs to evaluate the function of the right ventricle (RV )[28]. As a result, segmenting these two chambers, either manually or auto- matically, can provide lots of handy information for people who are interested in performing such kind of studies.
Short axis[29] is the most common imaging plane to assess cardiac function. Using these images, segmentation of left and right ventricles (LV and RV respectively) followed by obtaining volumes, masses and ejection fraction lead to cardiac con- tractile function quantification. Other cardiac planes may also be used to localize the chambers and gain more insights regarding measurements.
1.2.2.1 Cardiac Imaging Planes
Generally, body structures are described in relation to the anatomical positions such as subject standing upright or facing the observer. This principle helps clini- cians localize the regions of interest with terms such as right/left, inferior/superior or medial/lateral. However, due to the specific characteristics of the heart, it is considered to be an exception for this rule. Cardiac planes are defined according to their axes[30]. Use of the cardiac axes enables standard points of reference to be maintained in these imaging modalities. Below, the principal imaging planes for cardiac structures are explained briefly (see Fig. 1.5).
Body Axes: Axial, sagittal and coronal views can be obtained for heart. The axial plane is commonly used for morphological studies and also for observing the relationships of the four cardiac chambers and the pericardium. Sagittal images can be used to study the relationships of the great vessels and ventricles. Finally,
coronal images are widely useful for assessment of the left ventricular outflow tract (LVOT ), the pulmonary veins and the left atrium.
Cardiac Axes: In order to get these images, first axial scout view at the level of the LV is acquired. Then from this image, a new plane running through the apex of the LV and the middle of the left atrioventricular (AV ) mitral valve is chosen. This plane is called vertical long-axis (VLA) plane. From this plane, a new plane which transects the LV apex and the middle of the mitral ring is defined. This plane is called horizontal long-axis (HLA) plane. Perpendicular planes to VLA and HLA planes then give short-axis (SA) plane. From SA plane, four-chamber view (4CH ) can be obtained at the level of mitral valves. 4CH view passes starting from the most superior mitral valve, to the inferior angle of the RV, usually through the mid-point of the interventricular septum. Finally, a true-SA plane can be prescribed off the 4CH view perpendicular to the interventricular septum.
Left side of the heart: From the SA plane, different planes for the left side of the heart can be obtained. For example, the VLA, HLA, and LV inflow /outflow can be acquired. LV inflow /outflow image is obtained using a plane that passes across the center of the aortic and mitral valves on the basal SA slice, or by performing a 3-point acquisition. In the three-point plane selection approach, the first point is located on the LV apex, the second in the center of the mitral valve, and the third in the center of the aortic valve. The LVOT view (oblique coronal orientation) can be prescribed by passing an imaging plane through and perpendicular to the aortic valve.
Right side of the heart: Two-chamber view, right ventricular outflow tract (RVOT ) and inflow/outflow plane of the RV can be acquired. The two-chamber view can be obtained by locating a plane through the RV apex and the mid-point of the tricuspid valve on the 4Ch view. RVOT is obtained by aligning a plane that passes through the main pulmonary artery (PA) and the RV inferiorly from a set of axial images. Inflow/outflow plane can be acquired using a 3-point plane.
The first point is placed on the tricuspid valve, the second on the RV apex, and the third on the pulmonary valve.
Figure 1.5: Some common cardiac MRI planes.
1.2.2.2 Overview of Segmentation Methods
Image segmentation[31] is the process of partitioning an image into different re- gions using objects boundaries. The partitioned image is then easier to analyze and contains useful information to process according to the specific needs of the application. More specifically, in medical images (X-Ray, CT, MRI, microscopy,
PET, SPECT, Endoscopy, etc), segmentation is the process of extracting bound- aries for objects or regions of interest using 2D or 3D images[32]. Regions of interest can be organs, cells, tissues or any other objects under investigation. The outcome of the segmentation is then contouring surrounding the objects which can be used to obtain quantitative and qualitative measurements for further diagnostic insights (see Fig. 1.6).
Figure 1.6: An example of medical image segmentation (using one of the segmentation tools developed in this project).
Cardiac segmentation process usually consists two main steps: (i) extracting ROI centered on the heart in order to not process the whole image and also for decreas- ing computational load (ii) segmenting the ventricles, walls and other regions of interest inside the heart chamber. For doing both tasks, there are lots of different implemented and under investigation methods. To name some, thresholding, edge- based, region-based, pixel-based, deformable models, and atlas guided. It is also possible to merge some of these techniques or use other segmentation techniques combined with listed ones to propose a new method[33].
There are different considerations for evaluating the goodness of segmentation methods. Firstly, some methods are just for segmenting one ventricle border (mostly LV because of its simplicity) whereas others try to delineate the contours
for both ventricles. Another critical factor is the level of prior knowledge fed to segmentation algorithms. This prior knowledge is in different forms. It can be some user interaction such as pointing out centers of cavities or some of the pixels in ventricle borders. It can also be in the form of some initial assumptions such as spatial or anatomical relationships between objects. Training datasets with man- ually generated segmentations is also a useful and reliable technique to enhance the accuracy of segmentation. Image-based, pixel-based and deformable models usually perform segmentation based on no or weak prior knowledge whereas, in Atlas guided and some other similar techniques, substantial prior knowledge is needed. Including or excluding papillary muscles and trabeculae structures also has some effects in the final quantification process. Moreover, automatic detec- tion of end systolic and end diastolic frames which are the most critical slices in quantification are handled differently using different approaches. Some of them need manual interference of users whereas others detect these phases automati- cally. Finally, regarding motion, various techniques also have different levels of considerations. Some of them do not use this information and propagate initial segmentation results over the whole cardiac cycle, whereas some other approaches use it for performing image registration and tracking the components of interest over different phases of the cardiac cycle[34].
Usually, for evaluating the accuracy of segmentation methods, some validation is performed against the ground truth (manual delineations of experts). This evaluation is in two ways: (i) comparing the contours using the mean perpendicular distance between them or similar metrics (ii) quantitative measurement according to some extracted information such as ventricular volumes, masses or ejection fraction followed by some correlation analysis such as linear regression and Bland- Altman[35].
1.2.2.3 Considerations and Challenges
Both manual and automatic approaches suffer from some limitations. Manual seg- mentation is a time and energy consuming process and intra and inter-observer
variability exist in the results[36]. Automatic segmentation is limited occasionally in mid-ventricular slices. Moreover, due to the complex geometry of RV, these ap- proaches usually are limited to LV. Finally, the necessity of user interaction before performing segmentation or during it, limits the performance of such methods[37].
1.3 List of Tools, Equipment and Software
To perform imaging, computational tool development, validation and statistical analysis, following sets of tools, equipment and software were used in this work.
Chapter 2:
• image acquisition: 7T Agilent Varian[38], 7T Bruker[39]
• pulse sequence development: VnmrJ 2.2B, Varian MR systems
• image quality assessment and validation: ImageJ[40], Segment Medviso[41]
• PC-MRI reconstruction: Mono C#[42]
• PC-MRI processing: Matlab R2017a[43], Segment Medviso[41]
• statistical analysis: Matlab R2017a, GraphPad Prism 7.0[44], R[45]
Chapter 3:
• image acquisition: 7T Agilent Varian, 3T Philips Ingenuity TF whole body PET/MR Imaging System[46]
• parametric relaxation time mapping tool development: OsiriX DI- COM viewer[47], ImageJ
• statistical analysis: Matlab R2017a, GraphPad Prism 7.0
Chapter 4:
• image acquisition: 3T Philips Ingenuity TF whole body PET/MR Imaging System
• cardiac segmentation and quantification: Osirix DICOM viewer, Seg- ment Medviso
• statistical analysis: Matlab R2017a, GraphPad Prism 7.0
1.4 Objectives
This project was completed under the umbrella of Next generation training in car- diovascular research and innovation (CardioNext) which was part of Marie Curie Initial Training Networks (ITN) Call: FP7-PEOPLE-2013-ITN. CardioNext sci- entific objectives were to combine basic and applied innovative research on various heart diseases by (i) generation of computational models of the heart to better un- derstand its structure and model that can eventually be targeted for therapy, and (ii) implementation of forefront new imaging technology and therapeutic strategies to help clinicians and researchers. Global analysis and computational modelling of the heart and developing new non-invasive imaging techniques for MR imaging modality were the primary focus of this project. These objectives were accom- plished by getting essential pieces of training in the development of the innovative scientific objectives by the combined effort of top-quality CNIC researchers with complementary expertise in cardiovascular research together with collaborating academic partners and, more importantly, private companies at the forefront of cardiovascular R&D.
The aim of the work described in this thesis was to develop both imaging and computational tools to acquire and process cardiovascular related MRI data, al- though some of them can be used for other applications. In that sense, the project was divided into three subtasks:
1. Fast, High Spatial and High Resolution Magnetic Resonance Imaging of Small Animal Vessels using a Phase Contrast based Pulse Sequence
2. Linear Curve Fitting of Relaxation Time Maps in Magnetic Resonance Imag- ing
3. Automatic 4D Cardiac Segmentation and Quantification to Evaluate Ven- tricular Volume and Function in short-axis cine MR Images
All of the substasks defined above was supposed to be well implemented, tested and validated. For this purpose:
• comprehensive research on problems background was performed.
• various datasets coming from different sources were used for test and valida- tion.
• several validation criteria were defined to ensure that proposed methods are accurate and reliable.
• experts knowledge (cardiologists, physicians, etc) were used in the feedback loops for optimization of the tools and the results.
This investigation represents an attempt to address several challenges in the area of cardiovascular imaging, both in the level of image acquisition and postprocessing.
The findings of this study can be used by other researchers and clinicians to accelerate processing needed in the CMR area.
1.5 Thesis Outline
This thesis is organized in five chapters. The present chapter explains the problem of interest, its biological and computational research background, motivation and objectives for the research and the possible challenges and considerations. Chapter 2 gives a comprehensive explanation regarding the task Fast, High Spatial and High Resolution Magnetic Resonance Imaging of Small Animal Vessels using a Phase Contrast based Pulse Sequence. Chapter 3 describes the proposed solution
and experimental setup and results for task Linear Curve Fitting of Relaxation Time Maps in Magnetic Resonance Imaging. Chapter4is dedicated on discussion about the methodologies and outcomes of the proposed method for task Automatic 4D Cardiac Segmentation and Quantification to Evaluate Ventricular Volume and Function in short-axis cine MR Images. Chapter 5 summarizes the achievements of the thesis and points out possible research issues to be addressed in the future by other researchers.
Fast, High Spatial and High
Resolution Magnetic Resonance Imaging of Small Animal Vessels using a Phase Contrast based
Pulse Sequence
2.1 Introduction
MRI was introduced as an imaging modality more than forty years ago[48]. How- ever, the concept of designing advanced pulse sequences is more recent[49]. The main reason behind this is that in the initial days of imaging using MRI de- vices, RF fields, main magnetic fields, and gradients were used in a predefined manner and approaches such as iterative back projection reconstruction and T1 discrimination utilizing RF intensity change were embedded as standard modules to the devices. Later on, it turned out that by using time-dependent RF-pulses and magnetic field gradients, more flexibility is achieved in terms of contrast and plane definitions[2].
24
There are a wide variety of considerations when one tries to design a pulse se- quence. Some of them are listed below:
1. Magnetization preparation pulse design: to optimize the use of pre- pulses (e.g. magnetization transfer, fat suppression, inversion-recovery pulses, etc) before the excitation.
2. Radiofrequency pulse design: to design a radiofrequency (RF) pulse with sufficient bandwidth or adapted pulse shape for exciting all the resonance frequencies and control of excitation profiles in the region to be imaged.
3. Gradient lobes design: to control the phase that results from spin during the application of gradients. Standard (readout, phase and slice encoding) gradients and additional ones such as rewinder, spoiler, motion correction gradient lobes are used for this purpose.
4. Signal acquisition and K-space sampling: to manage temporary image space acquisition in frequency domain.
5. Physiologic gating, triggering and monitoring: to enhance/control the data acquisition accuracy after synchronization with a desired physiologic event.
6. Image data sampling: to optimize k-space data sampling from readout domains.
7. Order of the events (timing design of the sequence): to optimize the sequence design in order to obtain particular image appearance in a target range of period.
The issues mentioned above and many more minor ones[50] are between the units which are addressed in pulse sequence programming. The developed pulse se- quence ideally addresses all the technical considerations related to all or a subset of these items, and the output is a pulse sequence which can be embedded in the MRI machines and be used for different imaging or spectroscopic applications.
The main objective of this work was to develop and validate a Phase Contrast (PC ) based pulse sequence for performing high spatial, high temporal and fast imaging of small vessels in small animal. This sequence was not included in the list of software available for an Agilent-Varian system at the time this thesis was commenced (Agilent also left MRI business, so it was the only way to do this without such systems). To be more specific, the target pulse sequence had to meet these criteria:
1. Phase Contrast based (PC ): to image flow inside the blood vessels or to track spin motion, the sequence had to follow the basic idea of using flow or velocity encoding gradients to image moving magnetization[51].
2. Fast imaging: Repetition Time (TR) had to be decreased as much as possible (hence the total time needed) without losing remarkable accuracy and signal intensity in the resulting images[52] within the hardware limits.
3. High Spatial and Temporal resolution: refer to the definition of imaging limits with respect to time and space[9]. Spatial resolution generally determines the sharpness of the image and its level of captured details[10]. Temporal resolution is indicated by the interval between consecutive images.
Since at the date of doing this study, there were no other similar pulse sequences capable of acquiring such images in the MRI system we were using initially, val- idation of the proposed sequence had to be accomplished using an initial study.
Therefore, the results were validated using (i) image quality metrics analysis and (ii) by comparing the obtained results of the common measurements such as Flow, Peak Velocity and Pulse Wave Velocity with reported values in the literature.
Therefore, the whole pipeline consisted these tasks:
1. Pulse sequence development: A gradient echo based pulse sequence with flow compensation was implemented.
2. Image acquisition: The developed pulse sequence was used for acquiring raw small animal images.
3. Image reconstruction: Magnitude and phase images were reconstructed from raw images.
4. Sequence validation: The areas of interest (arterial lumen) were seg- mented followed by analysis in extracted regions. The quality of images in terms of SNR was assessed, as well.
2.2 Materials and Methods
2.2.1 Animal Model
Animal experiments were conducted in our institutional animal facility. All animal procedures were approved by our Institutional animal care and use committee (IACUC), and local authorities (Diputaci´on Foral de Guipuzcoa, Spain).
The study was performed using an established model of PAH in mice that is gen- erated by hypoxia exposure combined with Semaxanib (SU5416) intraperitoneal administration. Healthy normoxic mice were used as controls.
Eight-week-old male C57BL/6j mice were exposed to normobaric hypoxia (10% of oxygen) for 3 weeks and were only removed from the chamber once per week for the administration of intraperitoneal injections of the VEGF inhibitor. SU5416 was suspended in a vehicle containing carboxymethyl cellulose (CMC) (0.5%[w/v]
CMC sodium, 0.9%[w/v] sodium chloride, 0.4%[v/v] polysorbate 80, 0.9%[v/v]
benzyl alcohol in deionized water) and injected at 20mg/kg. Normoxic mice trated with IP injection of that vehicle were maintained in a room with normal oxygen levels.
2.2.2 Magnetic Resonance Imaging Protocol
Initial tests, sequence programming and images acquisitions were performed in a 7 Tesla Agilent-Varian MRI system[38]. Final experiments were performed in a Bruker 7T system, where the group was established at the end of the thesis.
In order to ensure that for validation of the sequence (see section 2.2.4), different type of subjects (healthy and with conditions) with different respiratory conditions will be used, in total fourteen mice divided into two sub-groups were used:
1. Hypoxia group: This group consisted six C57BL/6 mice (20-24 g) with hypoxia for three weeks. The experiments were performed exactly at 3 weeks after hypoxia exposure.
2. Healthy group: This group consisted four C57BL/6 male mice (20-24 g) without any health condition. Acquisitions for this group were performed during three respiratory conditions, i.e free breathing, inspiration, end of expiration.
Imaging was achieved with a 40mm inner diameter volume-coil working in quadra- ture for both signal transmission and reception. Animal preparation was started by inducing anaesthesia with 4% isoflurane (reduced to 1-2% during image ac- quisition), in a 30/70% mixture of O2/N2 a carrier gas. Animals were prevented from hypothermia with the use of a heater system with warm air controlled by fiber optic (SA Instruments). A pressure sensitive cushion was placed under the animal for respiratory synchronization and ECG electrodes were connected subcu- taneously to the front legs and left rear leg for cardiac synchronization. To ensure animal welfare and experimental conditions, temperature, respiration and heart rate were continuously monitored while they remain in the MRI magnet, using a SAII M1030 system (Small animal instruments, Stony Brook, NY, USA). This system was also used to synchronize image acquisition with the respiration and cycle cardiac of the animal.
For all subjects, the protocol focused on the exact and reproducible acquisition of the pulmonary artery (1 mm internal diameter approximately) at two points which are main trunk of the pulmonary artery (called MPA after this) and at the right ventricle outflow tract (called RPA after this). The selected regions were considered to be challenging ones in small animals due to the geometrical features (small vessel size, orientation, etc).
Our slices’ selection coincides with the most repeated localization of regions of interest found in the bibliography. Firstly, a four-chamber heart cine image was acquired in which the pulmonary artery could be clearly delimited below the aortic arch (see red dot in Fig. 2.1). On this plane, a perpendicular slice was selected to cut the pulmonary artery. The slice thickness was selected as 1.1 mm for both rats and mice so that we could obtain a sufficient signal / noise ratio. The result of this new selection is shown in Fig. 2.2.
Figure 2.1: Planning of the first 1.1 mm cut for reproducible selection of the pulmonary artery in both rat and mouse. The image also shows the position of the cut in a dotted line. It is done on a typical 4-chamber plane of the heart, parallel to the base of the heart and passing through the pulmonary artery.
These two steps allowed a reproducible selection of the pulmonary artery at the main trunk of the PA, before the first bifurcation. The second image was positioned on the right branch after the first bifurcation, just below the ascending aorta (see Fig. 2.3). The velocity or flow maps (see Fig. 2.4) were then obtained in two dimensions (the sequence was ready for 4D).
Figure 2.2: Image resulting from region planning shown in Fig. 2.1. The image shows the section of the main trunk of the pulmonary artery and the first branch. The red arrow indicates the first bifurcation of the pulmonary artery.
Figure 2.3: Localization of the plane perpendicular to the right branch of the pulmonary artery, immediately after the first bifurcation positioned below the
ascending aorta (arrow).
Data was acquired with dual cardiac and respiratory synchronization with 10, 15 or 20 phases of the cardiac cycle (depending on the type of analysis) at the end of expiration with minimum echo and repetition parameters. VENC (velocity of encoding measured in cm/sec) parameter had to be chosen in such a way to cover the whole range of velocities within the vessel of interest (minimum to maximum).
Usually, the chosen value for this parameter was slightly higher than the expected peak vessel velocity to avoid flow-related aliasing. Alternatively, an unwrapping algorithm could have been applied in the preprocessing stages[53]. In our case, we have tested VENC values of 50, 120 and 200cm/s. Sequence parameters for the
Figure 2.4: Flow maps obtained from the planes selected in Fig. 2.2 and Fig. 2.3. The image corresponds to one (out of the 10, 15 or 20 phases used) snapshot of the cardiac cycle. The arrow points to the pulmonary artery near the RV output tract (left) and on the first right branch after bifurcation (right
inside panel).
measurements are shown in Table 2.1.
Following this protocol, at the end of each acquisition, for original Agilent-Varian based sequence, the dataset of each subject, contained coupled magnitude and phase images for slice positions and selected number of cardiac frames. For Bruker based acquisitions, coupled magnitude images were acquired seperately in the same geometrical position.
2.2.3 Developed Pulse Sequence (Gradient Echo with Multi Slice and Flow Compensation)
The developed pulse sequence for an Agilent-Varian system was used for acquiring flow images. This sequence is considered to be a two-dimensional cine PC-MRI [51]
sequence and using it, qualitative and quantitative information such as blood velocity, flow or kinetic energy can be obtained.
Table 2.1: Image Acquisition parameters.
Parameter Value
Magnetic Field Strength 7.05T Number of Phase Encoding Steps 96
Matrix 128*128
Bandwidth 500.8Hz
Slice Thickness 1.1mm
Flip Angle 35◦
TE (minimum) 2.6ms
TR (minimum) 5.49ms
TE/TR (temporal resolution) 0.47
VENC 50, 120, 200cm/s (depending on type
of analysis)
Scan Time (inspiration) 40min (double gated) Scan Time (end expiration) 12min (double gated) Scan Time (no gating) 3.5min
Directional Components of Velocity 1
Number of Slices 2
Number of Frames 10, 15, 20 (depending on type of analysis)
Number of Averages 10
In summary, for each slice location and each cardiac cycle phase, a pair of magni- tude and phase images were used. Magnitude images were adopted for anatomical referencing whereas the phase images were used for extracting needed measure- ments.
The proposed technique has different building blocks:
1. Gradient Echo module: for obtaining magnitude anatomical images and as a base for developing PC related features.
2. Flow Encoding, Gradient Moment Nulling (Flow Compensation) and Phase Difference Reconstruction modules: for obtaining phase images for each slice position and cardiac cycle phase.
3. Cardiac Triggering module: to minimize the effects arising from cardiac motion and to obtain the magnitude and phase images during the whole cardiac cycle.