Multimodal Mitral Valve Segmentation: Mesh-Based 3D and Depth-Transfer Learning
Valvular heart disease affects a growing share of the population, with mitral valve regurgitation alone reaching a prevalence of 6.45% by age 90. Accurate mitral valve segmentation across imaging modalities is essential for diagnosis, intervention planning, and intraoperative guidance, yet robust automated tools remain scarce. We address this gap in the context of the Challenge "Mitral Valve Anatomy Analysis Using Multimodal Imaging Data" proposing modality-specific strategies for 3D transesophageal echocardiography (TEE), cardiac computed tomography (CT), and endoscopy. For TEE and CT, we introduce a mesh-based approach that better captures the thin, continuous leaflet geometry than voxel masks while preserving anatomical topology. Although voxel-based metrics rank our method in the lower third of the leaderboard, we show this largely reflects conversion artifacts, with competitive performance when compared to mesh-based literature. For endoscopy, we demonstrate that the transfer of depth prediction as an auxiliary task improves segmentation accuracy, achieving top third performance with a real-time capable architecture suitable for intraoperative use.
