Evidential Segmentation of Iliac Stenoses: A Case Study on Modeling Expert Variability

Abstract

The delineation of stenosis extent in digital subtraction angiography (DSA) images is inherently ambiguous and exhibits substantial inter-observer variability, which limits the suitability of deterministic segmentation approaches. To address this, we propose an uncertainty-aware segmentation framework based on evidential deep learning (EDL) that explicitly models annotation uncertainty from multiple expert segmentations. The proposed EDL U-Net is trained on ambiguously labeled DSA images using a tailored Beta-Binomial likelihood and produces a pixel-wise Beta distribution, enabling joint learning of segmentation and data uncertainty. The method is evaluated in a leave-one-out cross-validation on a dataset of iliac artery stenoses and compared to standard U-Net baselines. The EDL U-Net achieves a soft-Dice score of 0.83 ± 0.13, outperforming the baselines, while the learned uncertainty correlates with inter-observer variability as reflected by a low Brier score of 0.027±0.015 and an AURC of 0.040±0.040. These results demonstrate that explicitly modeling annotator variability via a Beta-Binomial evidential formulation leads to improved segmentation performance and principled uncertainty estimation in the presence of ambiguous annotations.