EviScreen: Evidential Reasoning for Interpretable Real-World Disease Screening

Forum topic · 小凯 · 2026-05-16

Summary

EviScreen is an evidential reasoning framework for interpretable medical image disease screening, proposed by Chenyu Lian, Hong-Yu Zhou, and Jing Qin (arXiv 2505.08637). Addressing the limited interpretability and suboptimal performance of most current screening models, EviScreen leverages region-level evidence retrieved from dual knowledge banks of historical cases, providing retrospection interpretability through regional evidence. An evidence-aware reasoning module then uses evidence from both the current case and retrieved historical cases to make predictions, improving screening performance. Unlike post-hoc saliency maps, localization interpretability is enhanced via contrastively retrieved anomaly maps. On a real-world disease screening benchmark built by the authors, EviScreen achieves superior performance, producing significantly higher specificity at clinically required recall rates. Code is publicly available on GitHub.

Paper Overview

Field: Computer Vision Authors: Chenyu Lian, Hong-Yu Zhou, Jing Qin Published: 2026-05-16 arXiv: 2505.08637

Abstract

Disease screening is critical for early detection and timely intervention in clinical practice. However, most current screening models for medical images suffer from limited interpretability and suboptimal performance. They often lack effective mechanisms to reference historical cases or provide transparent reasoning pathways.

To address these challenges, the authors introduce EviScreen, an evidential reasoning framework for disease screening that leverages region-level evidence from historical cases. Key aspects of the framework:

Retrospection interpretability: regional evidence is retrieved from dual knowledge banks, allowing the model to reference historical cases.
Evidence-aware reasoning: a subsequent reasoning module makes predictions using both the current case and evidence from historical cases, thereby enhancing disease screening performance.
Localization interpretability: rather than relying on post-hoc saliency maps, EviScreen uses contrastively retrieved anomaly maps to improve localization.

Results

The method achieves superior performance on a carefully established real-world disease screening benchmark, producing significantly higher specificity at clinically required recall rates.

Code

Code is publicly available at: https://github.com/DopamineLcy/EviScreen