Boundary delineation in prostate MRI is often unreliable in ambiguous regions, and standard overlap-based loss functions can miss small surface deviations. The authors propose Contrastive Uncertainty Learning (CUL), a supervised framework that turns predictive uncertainty into a training signal rather than treating it only as a post-hoc quality score.

CUL first uses Monte Carlo Dropout to estimate per-voxel uncertainty. Features with low uncertainty and correct predictions are pulled together as stable voxels, while high-uncertainty or misclassified features are actively pushed apart. This restructuring aims to keep the embedding space well separated at difficult boundaries. The authors tested CUL on a 589-scan dataset with a Vision Transformer and UperNet decoder, segmenting eight lower-pelvic structures.

Average Dice performance was essentially unchanged (83.38% vs 83.35%), but surface distance decreased significantly for six of eight structures, and a hard-case analysis found an 18.8% improved-versus-degraded ratio — a 22× asymmetry. Ablation experiments showed that removing uncertainty-guided sampling eliminated improvements for the transition zone, while swapping the contrastive objective for a boundary loss removed bladder improvements. The authors argue that both mechanisms are necessary and that CUL reduces the risk of extreme errors in safety-critical medical segmentation. The work appears in the UNSURE workshop at MICCAI 2026.