Sinkhorn-Triplet Loss Improves Multi-Organ Segmentation with Partial Supervision
September 18, 2026
A two-stage framework uses learnable organ prototypes and Sinkhorn-triplet loss to align feature distributions in partially annotated datasets. This method enforces organ-wise feature consistency across domain shifts, matching state-of-the-art performance on the BTCV dataset without requiring full annotations.
HOW THIS AFFECTS YOU
●
researcherYou can leverage Sinkhorn-triplet loss to enforce latent embedding consistency across heterogeneous imaging sources.
●
healthThis enables more accurate medical imaging segmentation when clinical datasets have incomplete manual annotations.