UC San Diego

Cardiac magnetic resonance (CMR) image segmentation has been a crucial tool for medical professionals to diagnose cardiovascular diseases (CVDs), which are the leading causes of death throughout the world. Segmenting CMR images is very time consuming and increases the cost of CVD diagnoses and treatment, making them inaccessible to many. Automated CMR image segmentation models strive to lower the cost of CVD diagnosis, but such models must be efficient and accurate in such failure-sensitive domains as human medicine. This thesis proposes to apply γ-Net, a recurrent extension of the popular U-Net, to automatically perform high-quality CMR image segmentation. γ-Net is a recent development by Linsley et al. of Brown University, and has exhibited the ability to outperform U-Net on very small datasets, which is beneficial given the very limited amount of patient CMR data available to the scientific community. γ-Net leverages biological principles backed by anatomical evidence as well as attention mechanisms in order to achieve its high efficiency.

In this thesis, we examine the following topics: (a) γ-Net’s resilience to smaller training set sizes, which is cruicial when little patient data is available; (b) resilience to variation in training and validation data, which is shown to significantly degrade performance in state-of-the- art models; and (c) the ability to transfer to new datasets with minimal fine tuning, which saves training cost for practical applications. We have found that (a) γ-Net significantly outperforms an equivalent U-Net in validation performance when trained using a reduced training set; (b) γ-Net is much more resilient to input variations than U-Net; and (c) γ-Net generalizes to new datasets better than comparable U-Nets.