Uncertainty-Supervised Interpretable and Robust Evidential Segmentation

SURE: Supervised UnceRtainty Estimation — making uncertainty more certain.

This repository contains the official implementation of the MICCAI 2025 paper:
"Uncertainty-Supervised Interpretable and Robust Evidential Segmentation".

🧠 Overview

In this work, we propose a novel uncertainty supervision framework guided by human-intuitive principles. Instead of treating uncertainty as a by-product of prediction, we explicitly supervise the uncertainty using interpretable patterns derived from human reasoning. Additionally, we introduce new evaluation metrics (UCC and UR) to evaluate the interpretability and robustness of model uncertainty.

---

📊 Metrics: UCC and UR

The proposed metrics are implemented in utilities/count_pixels.py, with the following functions:

• count_pixels_d
• count_pixels_d_chunk
• count_pixels_mu
• count_pixels_grad
• count_corr_mu

These functions are used to calculate the Uncertainty Correlation Coefficient (UCC) and Uncertainty Ratio (UR) metrics.

📌 Input Requirements

All input tensors for these functions — such as uncertainty, gradient, and distance maps — must be flattened 1D tensors of shape [N, 1].