Answers are grounded in this paper and cite their sources.
Data augmentation is used routinely to train segmentation networks, yet its effect on the reliability of their confidence is poorly understood. We trained a U-Net for dermoscopic lesion segmentation on ISIC-2016 under a baseline and four augmentation pipelines (geometric, intensity, mixing, and occlusion), each with three training seeds and a held-out validation split, and tested all fifteen models on ISIC-2016 and the external PH2 dataset. Uncertainty was estimated with Monte Carlo Dropout, and calibration was measured with the expected calibration error (ECE) and Brier score over all pixels and within a ± 5 -pixel band around the lesion contour. All comparisons used a hierarchical bootstrap over seeds and images. Dice did not differ significantly from the baseline for any pipeline (0.883–0.895), and global ECE did not differ significantly. Boundary ECE, however, was ten times larger than global ECE. On PH2, mixing augmentation reduced it from 0.257 to 0.209, significantly outperforming the baseline, intensity, and occlusion pipelines, and tempered the over-prediction of lesion probability that all models showed under domain shift. Mixing also produced the highest entropy but the lowest variance, so mean uncertainty is not a proxy for reliability, and calibration rankings changed between seeds. The effects of augmentation on reliability therefore become visible only with boundary-aware, multi-seed evaluation.
No takes yet. Share an insight, caveat, or question.
Banakhede et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: