Title: ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements

Abstract: To mitigate performance degradation caused by variations in imaging devices and clinical protocols across different domains, generalized medical image segmentation is essential. The Whitening Transform-based Probabilistic Shape Regularization Extractor (WT-PSE), previously published in IEEE Transactions on Medical Imaging (2024), tackles this issue by leveraging feature decorrelation and knowledge distillation based on the Wasserstein distance to ensure robust cross-domain segmentation. This paper conducts a systematic evaluation of enhancements to the WT-PSE learning framework. We identify four specific limitations in the original implementation: a restricted set of training augmentations that does not adequately mimic real-world scanner variations; a dependence on per-pixel binary cross-entropy loss, which is vulnerable to edge noise; the lack of a scheduled loss weighting mechanism, potentially leading to instability during early training phases; and the absence of ablation switches necessary for rigorous scientific comparison. In response, we introduce four targeted improvements: (1) domain-adaptive augmentation techniques, including random erasing, gamma correction, and salt-and-pepper noise injection; (2) a combined BCE and Dice loss function to enhance edge-aware segmentation in noisy environments; (3) a curriculum-based strategy for scheduling Dice weights; and (4) command-line control flags to facilitate systematic ablation studies. Our experiments on the fundus optic disc segmentation benchmark show that the refined pipeline achieves a final epoch optic-disc Dice score of 0.956 and an ASD score of 13.31, surpassing the baseline’s epoch-5 Dice score of 0.939. These findings suggest that refining the training process yields consistent performance benefits without requiring alterations to the core WT-PSE architecture.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC