Machine Learning Enabled Cardiac MRI Analysis for Congenital Heart Disease

Congenital heart disease (CHD) is among the most common birth defects, affecting approximately 1% of newborns. Cardiovascular magnetic resonance imaging (CMR) is considered the gold standard for detailed visualisation of cardiac anatomy and assessment of function in these patients. However, current evaluation of cardiac function relies heavily on manual post-processing, which is labour-intensive, time-consuming, and prone to variability. Recent developments in machine learning offer promising solutions for automating and accelerating CMR image analysis. This thesis presents three studies that investigate the application of machine learning techniques to key CMR analysis tasks in patients with CHD, with a focus on those with single-ventricle physiology within the FORCE registry. First, an automated pipeline was developed to identify, locate, and segment ventricles in short-axis CMR images from over 5,500 exams within the FORCE registry. The pipeline used these segmentations to calculate ventricular function metrics, including volume, mass, and ejection fraction. Second, a framework was designed to automatically segment five major blood vessels from all phase-contrast CMR scans in the FORCE registry. Flow curves derived from these segmentations were then clustered to identify distinct flow phenotypes and related these to adverse outcomes. Finally, a model combining image and graph convolutions was created to generate patient-specific pulmonary artery meshes and predict associated computational fluid dynamics flow fields. These methods enable automated, consistent CMR processing to support large-scale studies, inform clinical decisions, and improve treatment guidelines and outcomes for CHD patients.