Rel-UNet: Reliable Tumor Segmentation via Uncertainty Quantification in nnU-Net

Abstract

Accurate and reliable tumor segmentation is essential in medical imaging analysis for improving diagnosis, treatment planning, and monitoring. However, existing segmentation models often lack robust mechanisms for quantifying the uncertainty associated with their predictions, which is essential for informed clinical decision-making. This study presents a novel approach for uncertainty quantification in kidney tumor segmentation using deep learning, specifically leveraging multiple local minima during training. Our method generates uncertainty maps without modifying the original model architecture or requiring extensive computational resources. We evaluated our approach on the KiTS23 dataset, where our approach effectively identified ambiguous regions faster and with lower uncertainty scores in contrast to previous approaches. The generated uncertainty maps provide critical insights into model confidence, ultimately enhancing the reliability of the segmentation with the potential to support more accurate medical diagnoses. The computational efficiency and model-agnostic design of the proposed approach allows adaptation without architectural changes, enabling use across various segmentation models.