Unsupervised domain adaptation for multi-view echocardiographic segmentation via gaussian contrast and masked consistency learning

In clinical practice, cardiologists require multi-view echocardiographic segmentation for accurate cardiovascular disease diagnosis. However, segmenting echocardiograms from multiple views presents significant challenges, including structural variations, a lack of manual annotations, and inconsistent acquisition methods. This study proposes an unsupervised domain adaptive approach based on semantic contrastive learning to overcome these challenges. Specifically, we apply a teacher-student architecture to knowledge transfer, implementing semantic alignment and pseudo-label generation to handle cross-domain semantic variations through two learning strategies. In knowledge transfer, Gaussian Projection Contrastive Learning enhances cross-domain category discrimination by establishing intrinsic connections between cardiac features across different viewing angles. For pseudo-label generation, a masked uncertainty-aware consistency learning strategy enforces consistency between the target view's cardiac appearance and the generated pseudo-labels, leveraging contextual reasoning to predict masked regions. Extensive evaluations were conducted on echocardiographic datasets comprising 688 patients from both internal and external sources. We designed not only inter-view transfer experiments to validate the performance of our method but also cross-dataset transfer experiments to further verify its generalizability across diverse adaptation scenarios.