Contrastive Boundary-Aware Learning for Unsupervised Cross-Modality Whole Heart Segmentation

Whole heart segmentation plays a crucial role in the diagnosis of cardiovascular diseases and in treatment planning. Though many existing works achieve promising results, challenges remain due to domain discrepancies, scarcity of annotated data, and the complex anatomy of the heart. Unsupervised Domain Adaptation (UDA) has emerged as a promising solution to the scarcity of annotated data by transferring knowledge from labeled to unlabeled modalities. Many existing domain adaptation methods address the problems of domain distribution gaps through adversarial training and often generate erroneous results for small cardiac structures like myocardium. This remains a significant challenge due to insufficient boundary preservation and feature misalignment. In this work, we propose Contrastive Learning (CL) for feature alignment across Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) modalities without relying on global prototypes, especially for smaller and more complex regions like the myocardium. The integration of both dice and boundary-aware losses is employed to maximize overlap and to penalize discrepancies at the boundaries. This approach also enhances the precision of the boundaries. A substantial set of experiments was conducted on the Multi-Modality Whole Heart Segmentation (MM-WHS) dataset. The experimental results demonstrate significant improvements in segmentation accuracy, particularly in challenging regions such as myocardium. The experimental results yielded a mean Dice coefficient of 75.3% and an Average Symmetric Surface Distance (ASSD) of 2.7 mm, outperforming existing methods.