Evidential Reasoning Advances Interpretable Real-World Disease Screening

Problem
This paper addresses the gap in interpretability and performance of existing disease screening models for medical images. Current models often fail to provide transparent reasoning pathways and do not effectively utilize historical case data, which limits their clinical applicability. The authors propose EviScreen, an evidential reasoning framework designed to enhance both interpretability and predictive performance in real-world disease screening scenarios. This work is presented as a preprint and has not yet undergone peer review.

Method
EviScreen employs a dual knowledge bank system to retrieve region-level evidence from historical cases, which is integrated into a reasoning module that combines current case data with historical evidence. The architecture consists of two main components: the evidence retrieval mechanism and the evidence-aware reasoning module. The evidence retrieval mechanism utilizes contrastive retrieval to generate abnormality maps, which serve as a basis for localization interpretability. The reasoning module processes this evidence to make predictions, thereby improving the model’s performance. The authors do not disclose specific details regarding the training compute or the exact architecture used, but they emphasize the importance of the evidential mechanism in enhancing interpretability and performance.

Results
EviScreen demonstrates superior performance on established benchmarks for disease screening, achieving a notable increase in specificity while maintaining clinical-level recall. The authors report that EviScreen outperforms traditional models, although specific baseline models and quantitative metrics are not detailed in the abstract. The results indicate a significant improvement in the interpretability of predictions, as evidenced by the use of abnormality maps, which provide clearer insights into the model’s decision-making process compared to conventional saliency maps.

Limitations
The authors acknowledge that while EviScreen improves interpretability and performance, it may still be limited by the quality and representativeness of the historical case data used in the dual knowledge banks. Additionally, the reliance on contrastive retrieval for abnormality map generation may introduce biases if the historical data is not sufficiently diverse. The paper does not address potential computational overhead associated with the evidential reasoning framework, which could impact scalability in real-world applications.

Why it matters
The introduction of EviScreen has significant implications for the field of medical imaging and disease screening. By enhancing interpretability through evidential reasoning, the framework could facilitate clinician trust and adoption of AI-driven diagnostic tools. The ability to reference historical cases may also lead to more accurate and contextually relevant predictions, ultimately improving patient outcomes. This work paves the way for future research into integrating historical data into machine learning models, potentially transforming how disease screening is approached in clinical settings.

Authors: Chenyu Lian, Hong-Yu Zhou, Jing Qin
Source: arXiv:2605.15171
URL: https://arxiv.org/abs/2605.15171v1