Evidential Reasoning Advances Interpretable Real-World Disease Screening

Conference: ICML 2026
arXiv: 2605.15171
Code: https://github.com/DopamineLcy/EviScreen (available)
Area: Medical Imaging / Interpretable AI / Anomaly Detection
Keywords: Dual Knowledge Bank, Evidential Reasoning, Coreset Memory Bank, Contrastive Retrieval, Clinically-Oriented Evaluation

TL;DR

EviScreen employs a "normal + pathological" dual knowledge bank for region-level evidence retrieval, then performs evidential reasoning between the current case and retrieved evidence via cross-attention and self-attention. This provides both retrospective interpretability (which historical cases support the current decision) and localization interpretability (abnormality maps from contrastive retrieval). On four real-world external test sets, it achieves SOTA specificity at high recall.

Background & Motivation

Background: There are two mainstream approaches for disease screening in medical imaging: (a) Deviation-based prediction (unsupervised anomaly detection, e.g., PatchCore, SimpleNet) models only normal samples and flags deviations; (b) Direct prediction (binary classification, fully supervised), with post-hoc localization via Grad-CAM.

Limitations of Prior Work: (a) Pathological samples' rich information is not utilized, limiting performance on complex modalities (e.g., chest X-ray, dermoscopy); (b) Grad-CAM post-hoc maps have been shown to have poor localization quality and cannot explain "why this region appears abnormal"—lacking evidential reasoning. Prototype-based methods (e.g., ProtoPNet) use a fixed number of prototypes per class, which lack scalability and cannot cover the diversity of real clinical cases.

Key Challenge: Clinicians make decisions by "retrieving similar historical cases for comparison"; existing models either ignore historical cases or rely on "dozens of learned abstract prototypes"—disconnected from real diagnostic workflows.

Goal: (1) Design a screening framework that, like clinicians, can "retrieve similar region-level evidence from a scalable case bank"; (2) Establish clinically-oriented evaluation (external testing + specificity at high recall).

Key Insight: Reframe anomaly screening as a two-stage "retrieval + reasoning" process—extract region features using a foundation model, construct normal and pathological coreset knowledge banks; for each patch in the query image, perform \(k\)-NN retrieval in both banks, then use attention-based reasoning over these evidence tokens.

Core Idea: Using "normal vs pathological" dual knowledge banks for contrastive retrieval enables both localization (abnormality map \(\mathbf M = \text{ReLU}(\mathbf M_N - \mathbf M_P)\)) and reasoning (cross-attention integrates evidence into the query feature), turning "post-hoc Grad-CAM" into "built-in evidence flow".

Method

Overall Architecture

Two stages. Stage 1: Dual Knowledge Bank Construction: Split the training set into a "bank subset" \(\mathcal X^B_{N/P}\) and a "training subset" \(\mathcal X^R_{N/P}\); extract intermediate patch features using a frozen foundation model \(F_\theta\), then aggregate locally via \(\mathcal G_{agg}\) to form region feature sets \(\mathcal S_{N/P}\); apply greedy coreset subsampling (approximating the NP-hard problem of minimizing \(\max_m\min_n\|m-n\|_2\)) to obtain compact knowledge banks \(\mathcal K_N,\mathcal K_P\).

Stage 2: Reasoning: For a query image \(\mathbf x\), extract feature map \(\mathbf Z=\mathcal G_{agg}(F_\theta(\mathbf x))\in\mathbb R^{h\times w\times d}\); for each patch query \(\mathbf Z(i,j)\), perform \(k\)-NN in \(\mathcal K_N,\mathcal K_P\) to obtain evidence \(\mathbf E_N,\mathbf E_P\in\mathbb R^{h\times w\times k\times d}\); feed into the "evidence-aware reasoning" module with layered cross-attention (query ↔ evidence) + self-attention (inter-patch refinement); finally, concatenate the [CLS] tokens from both branches and send to MLP for prediction \(\hat y\). Also provides a training-free variant: directly pool scores from contrastive retrieval \(\mathbf M = \text{ReLU}(\mathbf M_N - \mathbf M_P)\).

Key Designs

1. Dual Coreset Knowledge Bank:

   • Function: Compactly and scalably represent the diverse morphologies of both "normal regions" and "pathological regions".
   • Mechanism: Extract patch features from \(\mathcal X^B_N,\mathcal X^B_P\) → local aggregation → greedy coreset subsampling to obtain \(\mathcal K_N,\mathcal K_P\). The optimization objective \(\mathcal K^*=\arg\min_{\mathcal K\subset\mathcal S}\max_{m\in\mathcal S}\min_{n\in\mathcal K}\|m-n\|_2\) is a classic NP-hard problem, approximated via iterative greedy selection for good coverage.
   • Design Motivation: Compared to ProtoPNet's "fixed \(K\) learned prototypes", coreset capacity can scale freely with data, covering the vast diversity of real clinical lesions; and since pathological images are inherently "normal + lesion mixed", only coreset (not class-specific prototypes) can accommodate such noisy data.

2. Evidence-Aware Reasoning:

   • Function: Truly integrate the \(k\) evidence vectors retrieved for each patch into the query feature, outputting new representations that consider "comparison with similar historical cases".
   • Mechanism: Each layer first performs cross-attention: \(\mathbf T^\ell_N(i,j)=\operatorname{softmax}\big(\mathbf Z^\ell_N(i,j)\mathbf E_N(i,j)^\top/\sqrt d\big)\mathbf E_N(i,j)\); then reshapes and applies inter-patch self-attention for refinement; normal branch \(\mathbf Z_N\) and pathological branch \(\mathbf Z_P\) run in parallel, final prediction \(\hat y=\text{MLP}([\mathbf Z_N^{\text{CLS}};\mathbf Z_P^{\text{CLS}}])\).
   • Design Motivation: Cross-attention explicitly writes "external evidence" into the feature map, ensuring each prediction is traceable to NN evidence (retrospective interpretability); dual branches preserve the two-dimensional signal of "similar to normal" vs "similar to pathological", better aligning with clinical intuition for differential diagnosis.

3. Contrastive Retrieval Abnormality Map (Training-Free Variant):

   • Function: Provides pixel-level abnormality localization without training any parameters, serving as a strong baseline and interpretability foundation.
   • Mechanism: For each patch, compute average \(k\)-NN distances to both banks \(\mathbf M_N,\mathbf M_P\in\mathbb R^{h\times w}\); the abnormality map is \(\mathbf M(i,j)=\text{ReLU}(\mathbf M_N(i,j)-\mathbf M_P(i,j))\)—only regions "far from normal but close to pathological" are highlighted. The final score is pooled from \(\mathbf M\).
   • Design Motivation: Pure PatchCore uses only \(\mathbf M_N\), mislabeling all rare normal variants; the "normal–pathological" difference filters out non-pathological deviations, yielding more focused localization; also provides an inherently different, endogenous localization map compared to Grad-CAM.

Loss & Training

Training: Only the "evidence-aware reasoning module" (cross/self-attention + MLP) is trained; the foundation model is frozen; loss is binary BCE. The training-free variant has no learnable parameters. \(k\), knowledge bank size, and number of cross-attention layers \(L\) are main hyperparameters.

Key Experimental Results

Main Results

10 public datasets, three major modalities (fundus, chest X-ray, dermoscopy), with focus on four external test sets: JSIEC, RIADD, CheXpert, Derm12345. Clinically-oriented metrics: AUROC, AP, Spe@95%R (specificity at 95% recall), Spe@99%R (specificity at 99% recall). Results (percent):

Metric	EviScreen	EviScreen-TF (training-free)	FM	PatchCore*	PatchCore	NFM-DRA	DRA	SCRD4AD	EDC	SimpleNet	CIPL
AUROC	98.06	96.76	95.84	94.96	92.12	95.53	92.53	94.88	79.12	73.73	94.83
AP	96.10	94.20	94.24	89.61	86.62	93.23	89.53	89.85	71.44	57.66	91.36
Spe@95%R	94.74	91.48	87.95	87.26	81.09	90.37	80.12	88.50	51.45	53.81	87.33
Spe@99%R	91.62	87.74	79.29	83.31	—	—	—	—	—	—	—

Ablation Study

Configuration	Key Effect	Notes
Full EviScreen	Best on all metrics	Dual bank + reasoning module
Remove pathological bank (only \(\mathcal K_N\))	Spe@95%R drops	Validates irreplaceability of dual bank
Use fixed prototypes (ProtoPNet style)	Performance drops	Coreset has larger capacity than fixed prototypes
Cross-attention only (no self)	Slight drop	Inter-patch refinement provides contextual consistency
Training-free variant	Outperforms PatchCore	Contrastive retrieval alone is a strong baseline

Key Findings

• Greater clinical metric gap: Compared to AUROC, this method's lead over baselines is more pronounced on Spe@95%R and Spe@99%R (e.g., Spe@99%R is 8.3 points higher than PatchCore*), highlighting the advantage of evidential reasoning in clinically critical high-recall regions.
• Training-free variant surpasses PatchCore: Contrastive retrieval alone outperforms traditional anomaly detection baselines, validating the effectiveness of the "dual bank + contrastive" approach.
• Scalability: Coreset size can scale linearly, facilitating continual addition of new cases—something prototype methods cannot achieve.

Highlights & Insights

• Dual-track interpretability: Provides both retrospection and localization explanations—far closer to radiologists' real diagnostic workflow than a single Grad-CAM heatmap.
• Clinically-oriented evaluation framework: 10 datasets + external testing + Spe@high-Recall as core, among the few truly deployment-oriented medical imaging evaluation designs; can serve as a community reference.
• Transferable "contrastive retrieval + attention fusion": This paradigm can be directly applied to other domains (industrial defect detection, satellite change detection) requiring "normal vs abnormal" discrimination.

Limitations & Future Work

• Dual bank construction requires sufficient and representative pathological samples; for rare or emerging diseases, "sparse pathological bank" may render contrastive retrieval ineffective.
• The foundation model is fully frozen, so performance ceiling is limited by it; continual replacement with medical domain foundation models can bring gains.
• Each patch performs \(k\)-NN during inference, causing significant latency for large banks; combining coreset growth strategies with approximate NN acceleration is a clear extension.
• Current reward/loss does not explicitly supervise "evidence–prediction consistency"; future work can add contrastive terms to ensure retrieved evidence truly guides prediction.

Related Work & Insights

• vs PatchCore / SimpleNet: Both use coreset memory banks, but only normal bank; EviScreen introduces a pathological bank for contrast, enabling more precise localization via contrastive retrieval.
• vs ProtoPNet / Prototype methods: Fixed prototype count is limited by preset classes; coreset capacity is flexible and covers a broader range.
• vs Grad-CAM post-hoc explanations: This work's explanations are "endogenous" (from retrieval + cross-attention), not reliant on gradient visualization tricks, yielding higher and more stable explanation quality.

Rating

• Novelty: ⭐⭐⭐⭐ The combination of dual coreset, contrastive retrieval, and dual-track interpretability is a novel and complete solution for medical screening tasks.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 10 datasets, 3 modalities, external testing + clinical metrics + comprehensive training-free ablation.
• Writing Quality: ⭐⭐⭐⭐ Clear structure with three key limitations and three main contributions; Figure 1's comparison of three paradigms is concise and effective.
• Value: ⭐⭐⭐⭐⭐ Provides a deployable clinically-oriented pipeline and evaluation protocol, contributing both methodology and benchmark to the medical imaging community.

Related Papers

• [CVPR 2026] XSeg: A Large-scale X-ray Contraband Segmentation Benchmark for Real-World Security Screening
• [ICML 2026] Marrying Generative Model of Healthcare Events with Digital Twin of Social Determinants of Health for Disease Reasoning
• [AAAI 2026] PulseMind: A Multi-Modal Medical Model for Real-World Clinical Diagnosis
• [AAAI 2026] Experience with Single Domain Generalization in Real World Medical Imaging Deployments
• [ACL 2026] OmniCompliance-100K: A Multi-Domain Rule-Grounded Real-World Safety Compliance Dataset