Evidential Reasoning Advances Interpretable Real-World Disease Screening

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, we introduce EviScreen, an evidential reasoning framework for disease screening that leverages region-level evidence from historical cases. The proposed EviScreen offers retrospection interpretability through regional evidence retrieved from dual knowledge banks. Using this evidential mechanism, the subsequent evidence-aware reasoning module makes predictions using both the current case and evidence from historical cases, thereby enhancing disease screening performance. Furthermore, rather than relying on post-hoc saliency maps, EviScreen enhances localization interpretability by leveraging abnormality maps derived from contrastive retrieval. Our method achieves superior performance on our carefully established benchmarks for real-world disease screening, yielding notably higher specificity at clinical-level recall. Code is publicly available at https://github.com/DopamineLcy/EviScreen.

• The paper introduces EviScreen, a dual knowledge bank framework that integrates historical evidence for improved interpretability and prediction in disease screening.
• It employs patch-wise evidence retrieval and cross-attention mechanisms, achieving superior AUROC, specificity, and clear separation rates across clinical benchmarks.
• The methodology enhances clinical trust with clear abnormality maps and demonstrated scalability, reducing ambiguous cases and supporting efficient deployment.

Evidential Reasoning for Interpretable Disease Screening: EviScreen

Introduction

Contemporary AI approaches for medical image-based disease screening are hampered by inadequate interpretability and limited ability to reference historical evidence, diverging from established clinical reasoning protocols. This paradigm often leads to suboptimal specificity and recall, and lack of transparency in decision-making processes. "Evidential Reasoning Advances Interpretable Real-World Disease Screening" (2605.15171) introduces EviScreen, an evidential reasoning framework that integrates region-level evidence from dual knowledge banks of historical normal and pathological cases. The method is designed to simultaneously improve prediction performance and interpretability for real-world clinical disease screening tasks. EviScreen is empirically validated through comprehensive benchmarking across ophthalmology, radiology, and dermatology modalities, deploying clinically relevant metrics.

Methodology

Dual Knowledge Bank Construction

EviScreen establishes two scalable coreset-based knowledge banks: one for normal cases, one for pathological. Patch-level features are extracted from a vision foundation model (ViT-based architectures, e.g., RETFound-Dinov2, CheXFound, PanDerm) and aggregated to form compact banks via greedy approximation based on Euclidean distance for redundancy minimization. These banks hold a rich representation of diverse normal and pathological regional features, addressing limitations of fixed prototype-based interpretable models.

Evidence Retrieval and Evidence-Aware Reasoning

For a query image, regional features are extracted and used to retrieve k-nearest neighbors from both banks. This enables patch-wise evidence retrieval, providing query-specific visual evidence. Prediction is generated via evidence-aware reasoning: a cross-attention mechanism incorporates retrieved evidence into the feature space, followed by self-attention for inter-patch refinement. Ultimately, MLP aggregates evidence-aware features for classification. This pipeline supports retrospection interpretability, as the model references historical visual analogs, and enhances localization interpretability via abnormality maps.

Contrastive Retrieval: Training-Free Variant

A training-free variant utilizes the dual banks for abnormality localization. Average distance maps to the normal and pathological banks are computed per patch; their contrast identifies abnormal regions, yielding abnormality maps with higher focus and clarity than deviation-based approaches (e.g., PatchCore).

Evaluation Framework and Benchmarking

The paper develops a clinically oriented evaluation protocol, prioritizing specificity at high recall to match practical requirements. Benchmarks span ten public datasets in three domains: JSIEC, RIADD, CheXpert, Derm12345, among others. Metrics include AUROC, Average Precision, Specificity at X% Recall (Spe@X%R), and Clear Separation Rate (CSR)—quantifying overlap in prediction score distributions.

Empirical Results and Claims

Numerical Performance

EviScreen consistently surpasses state-of-the-art (FM fine-tuning, PatchCore, SCRD4AD, SimpleNet, NFM-DRA, CIPL) across all test settings. For example, on JSIEC:

• AUROC: EviScreen 98.06%, FM 95.84%, PatchCore* 92.12%
• Spe@100%R: EviScreen 91.27%, FM 78.39%, PatchCore* 82.34%
• CSR: EviScreen 88.95%, FM 80.02%, PatchCore* 81.61%

Similar margins are maintained across RIADD (CSR: 54.38% vs FM 48.91%), CheXpert (CSR: 70.59% vs FM 57.41%), and Derm12345 (CSR: 77.92% vs FM 55.09%).

The method also achieves clear separation rates in pathological vs normal cases, indicating practical utility by reducing ambiguous predictions requiring manual review. Ablation analysis demonstrates that both evidence retrieval and evidence-aware reasoning are necessary for optimal performance; removal of either component reduces Spe@100%R and CSR by up to 38%.

Interpretability

EviScreen provides retrospection interpretability by retrieving relevant patches from the knowledge banks for each query region, mimicking clinician workflow. Abnormality maps generated via contrastive retrieval exhibit more spatial precision and focus compared to saliency maps and deviation maps. Quantitative reader studies corroborate superior Dice (0.66 vs PatchCore* 0.45) and IoU (0.50 vs PatchCore* 0.30) scores in localization.

Robustness and Scalability

Stress testing with contaminated pathological banks and explicit denoising show minimal impact on AUROC, Spe@99%R, validating robustness to noisy patch-level evidence. Scaling analysis demonstrates that increasing the knowledge bank size enhances clinically oriented metrics (AUROC, Spe@X%R, CSR).

Deployment Efficiency

Memory and latency analyses indicate efficient deployment: largest banks (1M vectors, 4GB) require ~0.6s retrieval.

Implications and Theoretical Impact

The work introduces a scalable, evidence-based reasoning paradigm that directly mirrors clinical decision approaches, thereby enhancing trustworthiness and transparency in medical AI. The dual knowledge bank mechanism resolves issues of limited prototype diversity in prior interpretable models and leverages region-level evidence for both prediction and interpretation, promising broader generalization and adaptability. By advancing clinically meaningful evaluation protocols, the paper lays groundwork for more rigorous benchmarking in medical screening AI.

Practically, the reduction in ambiguous cases (higher CSR) and improved specificity at high recall (Spe@X%R) imply lower rates of unnecessary follow-ups and missed diagnoses, offering direct benefit in real-world clinical workflows.

Future Directions

EviScreen's evidential reasoning pipeline is extensible to 3D imaging, broader modalities, and finer-grained screening tasks. Further research may integrate multi-modal data, enhance privacy-preserving mechanisms in historical evidence storage, and explore optimization strategies for large-scale deployment. This framework inspires a shift toward transparent, evidence-linked prediction pipelines in AI-driven diagnostics, possibly affecting future standard-of-care protocols.

Conclusion

"Evidential Reasoning Advances Interpretable Real-World Disease Screening" (2605.15171) introduces a dual knowledge bank-based evidential reasoning framework, markedly advancing both interpretability and clinical performance in disease screening tasks. The empirical evidence supports substantial gains in specificity at clinical-level recall, clear separation rates, and computational efficiency. This methodology represents a step toward interpretable, evidence-driven AI that aligns with clinical best practices, with notable potential for broad application and future refinement in medical AI research.