Uncertainty-Supervised Interpretable and Robust Evidential Segmentation

TL;DR

This work tackles unreliable and opaque uncertainty estimates in medical image segmentation. It introduces a self-supervised uncertainty supervision framework based on evidential deep learning (EDL) that ties uncertainty to boundary gradients and local noise via two losses, $L_{gu}$ and $L_{nu}$, plus an active-hard-sample mechanism, integrated into $L_{total}=L_{Seg}+\beta L_{gu}+\gamma L_{nu}$. The authors propose new interpretability metrics $UCC$ and $UR$ to quantify alignment between uncertainty and image structure/noise, and demonstrate on the ACDC and REFUGE datasets that the method achieves competitive segmentation performance while significantly improving uncertainty interpretability and robustness, especially under distribution shifts. Code is released at https://github.com/suiannaius/SURE.

Abstract

Uncertainty estimation has been widely studied in medical image segmentation as a tool to provide reliability, particularly in deep learning approaches. However, previous methods generally lack effective supervision in uncertainty estimation, leading to low interpretability and robustness of the predictions. In this work, we propose a self-supervised approach to guide the learning of uncertainty. Specifically, we introduce three principles about the relationships between the uncertainty and the image gradients around boundaries and noise. Based on these principles, two uncertainty supervision losses are designed. These losses enhance the alignment between model predictions and human interpretation. Accordingly, we introduce novel quantitative metrics for evaluating the interpretability and robustness of uncertainty. Experimental results demonstrate that compared to state-of-the-art approaches, the proposed method can achieve competitive segmentation performance and superior results in out-of-distribution (OOD) scenarios while significantly improving the interpretability and robustness of uncertainty estimation. Code is available via https://github.com/suiannaius/SURE.

Figures (3)

• Figure 1: The overview of our work. Based on EDL, our model estimates uncertainty from evidence, with gradient-based and noise-based supervision to enhance both interpretability and robustness of uncertainty.
• Figure 2: Robustness evaluation on ACDC dataset, different colors represent different methods. The values are the difference of scores under two different noise levels, e.g.,$\triangle DSC(0.3,0.1) = DSC(\mu=0.3)-DSC(\mu=0.1)$.
• Figure 3: Illustration of prediction results and uncertainty maps of different methods. Red arrows/boxes highlight erroneous uncertainty estimations, while green indicates correct ones.