DistMedVL: Probabilistic Vision-Language Alignment for Uncertainty-Aware Medical Image Segmentation

A new probabilistic vision-language framework, DistMedVL, has been developed by researchers to enhance medical image segmentation by explicitly addressing uncertainty in cross-modal alignment. This framework tackles the shortcomings of current deterministic approaches that fail to account for aleatoric uncertainty from unclear boundaries and epistemic uncertainty due to insufficient training data, resulting in fragile performance during domain shifts. DistMedVL utilizes a lightweight Probabilistic Cross-Modal Adapter (PCM-Adapter) built on frozen encoders to represent uncertainty. It features two sequential components: a Mahalanobis Alignment Module (MAM) that represents textual tokens as Gaussian distributions and assesses patch-text compatibility, alongside a second module for progressive probabilistic alignment. This method seeks to improve robustness in clinical settings where uncertainty is common. The research can be found on arXiv under identifier 2608.05683.