Left Ventricular Myocardium Segmentation and Registration Using Weakly-Supervised Learning Techniques

Ischemic cardiomyopathy (ICM) is a condition caused by myocardial ischemia and hypoxia due to coronary artery atherosclerosis, leading to heart dysfunction and symptoms like angina and heart failure. As a major cause of heart failure, ICM increases morbidity rates significantly. While angiography is commonly used to assess myocardial blood supply, it is invasive, costly, and requires expert interpretation. Non-invasive methods, such as deep learning (DL)-based segmentation of computed tomography angiography (CTA), show promise in providing accurate assessments with reduced reliance on invasive procedures. However, existing DL methods often struggle with generalization and require large annotated datasets. This study addresses these challenges by proposing a novel framework that combines the unsupervised Voxelmorph method, which uses U-Net and a spatial transformer network, with weakly supervised techniques for myocardial blood supply segmentation and registration. Instead of requiring full annotations for all blood supply regions, we reduce the annotation burden by using weak supervision, annotating only the left ventricular (LV) myocardium. This approach significantly lowers the need for large-scale labeled data while maintaining robust segmentation performance. For registration, the unsupervised Voxelmorph framework is employed, which ensures accurate alignment of myocardial regions while preserving topological consistency across datasets. The proposed method shows superior performance in terms of segmentation accuracy and registration precision, evaluated using metrics such as the Dice coefficient, registration error, and mean square error.