Integrating Uncertainty Into U-Net Robustness Evaluation Under Natural MRI Alterations: Application to Kidney Segmentation

The adoption of deep learning (DL) for medical image segmentation in clinical practice is limited by interpretability issues and sensitivity to out-of-distribution data. Robust models that maintain stable accuracy while offering reliable uncertainty estimates in the presence of data variations should be required. We propose a framework to evaluate the robustness of U-Net-based segmentation on a kidney MRI dataset, simulating common abdominal MRI distortions. Robustness is assessed using a novel metric that evaluates both the network’s accuracy stability and uncertainty reliability. The results show that, while segmentation accuracy remains stable across alterations, uncertainty is more sensitive to these changes. This suggests that capturing also uncertainty offers a more comprehensive assessment of DL models than traditional accuracy-focused frameworks.