Impact of task-informed data augmentation on U-Net-based kidney segmentation performance

Paper in proceeding, 2026

Deep learning-based kidney segmentation in magnetic resonance imaging (MRI) is sensitive to acquisition-related distribution variations, potentially compromising both accuracy and reliability. This study evaluates whether task-informed data augmentation can improve accuracy, uncertainty and robustness in a U-Net framework. Three physically plausible MRI perturbations were incorporated during training: affine transformations, bias field-induced intensity inhomogeneities, and Gaussian blurring. Performance was evaluated under controlled distortions during testing, by jointly analyzing Dice similarity and predictive uncertainty estimated via Monte Carlo Dropout. Robustness was quantified through accuracy degradation and an uncertainty-aware augmented robustness metric. We found that all data augmentation strategies outperformed the non-augmented baseline. The combination of affine and bias field transformations achieved the highest Dice and the best uncertainty-aware robustness across cohorts. These results demonstrate that task-informed augmentation improves both segmentation stability and uncertainty reliability under realistic distribution shifts, supporting more robust clinical implementation.