U-Net variants commonly encounter limitations due to overconfidence in predictions, impeding their clinical applicability. Quantifying model uncertainty accurately is vital, but obtaining sufficient and reliable evidence remains challenging. This paper introduces the \(\textbf{D}\) ual-Branch \(\textbf{Evi}\) dential Framework Fusing \(\textbf{H}\) ard \(\textbf{E}\) xample \(\textbf{M}\) ining for Abdominal Organ Segmentation (DEvi-HEM). It is a novel dual-branch framework integrating hard example mining (HEM) at region and pixel levels. By applying higher penalty weights to hard examples, HEM improves fine-grained prediction. The dual-branch structure enhances the model’s expressiveness by learning from both region-level and pixel-level representations. Furthermore, the introduction of dual-branch consistency learning and adversarial learning-based variational distributions captures the cognitive variability across branches. This ensures precise segmentation and reliable uncertainty estimation. DEvi-HEM improves segmentation performance, cuts computational cost, and outperforms uncertainty-based methods, with 3.292 GFLOPs on FLARE22 and 2.486 GFLOPs on Synapse.
