Since its invention in the 1970s, magnetic resonance imaging (MRI) has contributed greatly to our understanding of the human body in health and disease. MRI images anatomy and physiology with high spatiotemporal resolution and without ionizing radiation. Due to these factors, MRI is particularly well suited to studying the heart, and cardiac MRI is considered the clinical gold-standard for assessment of cardiac morphology, flow, and function. However, interpretation of cardiac MRI is highly dependent on image quality and often requires extensive manual annotation and visual analysis. Convolutional neural networks (CNNs), a form of deep learning and artificial intelligence, have potential to revolutionize medical imaging. Broadly, CNNs comprise a series of trainable weights called layers, which iteratively learn the features required to perform a given task. Currently, CNNs are being explored for a variety of computer vision tasks, such as classification, localization, and segmentation, but have untapped potential. Specifically, their ability to perform image synthesis is unknown.
Given these challenges in MRI and the untapped potential of CNNs, I asked: can we use deep learning to perform image synthesis for MRI? Using this question as the bedrock for my dissertation, I set out to solve progressively more challenging problems in cardiovascular MRI using CNNs, building towards the ultimate task of automatically quantifying cardiac function and biomechanics. In aim 1, I asked whether existing CNNs can enhance low-resolution cardiac images. That is, can CNNs perform image super-resolution of steady-state free precession (SSFP)? Specifically, I asked which CNN architectures are suitable for this task and how well they perform relative to conventional image upscaling methods.
In aim 2, I asked whether I could upgrade the CNN architectures from aim 1 to isolate and remove background signal from 4D Flow MRI. That is, can CNNs perform phase-error correction of 4D Flow MRI acquisitions via synthesis of the background static vector field? To achieve this, I asked what architectural modifications are necessary to infer these multi-component volumetric vector fields. I then compared CNN-based phase error correction with existing manual segmentation-based methods.
In aim 3, I asked whether I could further upgrade my phase-error correction CNN from aim 2 to predict intracardiac blood flow from videos of the beating heart. That is, can CNNs infer dynamic blood flow velocity fields from cardiac cine SSFP images? Specifically, I asked how I could incorporate spatiotemporal information and anatomical boundaries into this new architecture I call Triton-Net. I then measured the correlation between the synthesized flow fields and 4D Flow MRI measurements. Lastly, I asked whether I could use these flow values to detect left ventricular outflow obstruction.
Finally, in aim 4, I asked whether I could refine Triton-Net to evaluate local myocardial function. That is, could I add explicit physical constraints into Triton-Net to infer dynamic myocardial velocity and strain tensor fields from cardiac cine SSFP images? Realizing that myocardial contraction is periodic, I explored how I may encode net-zero displacement and strain constraints into the Triton-Net architecture, resulting in a heavily modified deep learning synthetic strain (DLSS) CNN. I then characterized DLSS strain in a healthy population and asked whether I could use DLSS strain to identify wall motion abnormalities in an ischemic heart disease population. Lastly, I compared DLSS classification performance against the consensus visual assessment of four cardiothoracic radiologist readers.
