EviScreen: Evidential Framework for Disease Screening

• EviScreen is an evidential reasoning framework that constructs dual region-level knowledge banks to facilitate binary screening of medical images.
• It integrates retrospection interpretability by retrieving historical normal and pathological evidence and localization interpretability via contrastive abnormality maps.
• The framework achieves high specificity at very high recall through foundation-model feature extraction, transformer-style reasoning, and efficient FAISS retrieval.

EviScreen is an evidential reasoning framework for real-world disease screening with medical images. It is organized around two coupled mechanisms: the construction of dual region-level knowledge banks from historical normal and pathological cases, and evidence-aware prediction for a new case through explicit retrieval of similar regions from those banks. In the formulation reported for ophthalmology, radiology, and dermatology, EviScreen performs binary screening of an image as normal or pathological, while also providing two forms of interpretability: “retrospection interpretability,” in which retrieved historical patches serve as explicit supporting or countervailing evidence, and localization interpretability, in which abnormality maps are derived from contrastive retrieval rather than post-hoc saliency (Lian et al., 14 May 2026).

1. Clinical setting and problem formulation

EviScreen is designed for disease screening in three medical imaging domains: color fundus photography in ophthalmology, chest X-rays in radiology, and dermoscopic images in dermatology (Lian et al., 14 May 2026). The task is binary disease screening, with image-level labels defined as normal versus pathological. In ophthalmology, normal denotes no detected fundus abnormalities and pathological denotes any fundus abnormality, including diabetic retinopathy, maculopathy, and CRVO. In radiology, normal denotes “No Finding,” while pathological denotes any abnormal finding excluding “Support Devices,” with evaluation on Atelectasis, Cardiomegaly, Consolidation, Edema, and Pleural Effusion. In dermatology, “normal” is operationalized as benign lesions and pathological as malignant lesions including MEL, LMM, SCC, and DFSP (Lian et al., 14 May 2026).

The framework is explicitly motivated by real-world constraints: noisy and heterogeneous data, severe class imbalance, distribution shift between training and external validation datasets, and the fact that pathological images contain both normal and abnormal regions while labels are usually only image-level (Lian et al., 14 May 2026). The paper positions these constraints against three mainstream paradigms. Deviation-based prediction can yield deviation maps but cannot fully exploit positive cases and does not provide explicit pathological references. Direct prediction can use both normal and pathological images but typically depends on post-hoc saliency methods such as Grad-CAM and Grad-CAM++, which are described as noisy and unreliable in medical imaging. Prototype-based interpretable models employ fixed prototype sets, which limits representational capacity in highly diverse pathology and does not scale naturally to a normal-versus-abnormal screening setting (Lian et al., 14 May 2026).

Formally, with image $x \in \mathbb{R}^{H \times W \times C}$ and label $y_x \in \{0,1\}$, EviScreen learns a screening function

$f : \mathbb{R}^{H \times W \times C} \to [0,1]$

that outputs a pathological score $\hat{y}$ while aiming for high specificity at very high recall, external generalization, and evidence-based interpretability (Lian et al., 14 May 2026).

Domain	Training data	External test data
Ophthalmology (CFP)	EDDFS, BRSET	JSIEC, RIADD
Radiology (CXR)	MIMIC-CXR	CheXpert
Dermatology (dermoscopy)	HAM10000, BCN20000, PAD-UFES-20	Derm12345

A defining feature of the framework is its clinically oriented evaluation. Rather than relying only on AUROC and AP, the study emphasizes specificity at very high recall thresholds and Clear Separation Rate, reflecting the operational requirement that screening systems should miss as few pathological cases as possible while minimizing false positives (Lian et al., 14 May 2026).

2. Dual knowledge banks

The first stage of EviScreen is offline construction of dual knowledge banks, one normal and one pathological, from historical cases (Lian et al., 14 May 2026). The normal and pathological training sets are partitioned as

$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$

where $X_N^R$ and $X_P^R$ are used to build the banks and $X_N^E$ and $X_P^E$ are reserved for training the evidential reasoning module (Lian et al., 14 May 2026).

A frozen vision foundation model $F_e$ is used to extract regional features. The modality-specific backbones are RETFound-Dinov2 for fundus images, CheXFound for chest X-rays, and PanDerm for dermoscopy (Lian et al., 14 May 2026). Intermediate features from ViT layers 7 and 17 are pooled and then passed through a locally aware aggregation function $y_x \in \{0,1\}$0, following the construction

$y_x \in \{0,1\}$1

$y_x \in \{0,1\}$2

Each aggregated representation is an $y_x \in \{0,1\}$3 map of patch-level features (Lian et al., 14 May 2026).

Because direct storage of all patch vectors would be redundant and computationally expensive, EviScreen applies coreset subsampling to both repositories:

$y_x \in \{0,1\}$4

$y_x \in \{0,1\}$5

The result is a normal knowledge bank $y_x \in \{0,1\}$6 and a pathological knowledge bank $y_x \in \{0,1\}$7, both stored in FAISS for fast nearest-neighbor retrieval (Lian et al., 14 May 2026).

These banks are region-level rather than image-level repositories. Each stored vector corresponds to local appearance structure, such as vessels and optic-disc patterns in fundus photography, lung-field and pleural patterns in chest radiography, or lesion texture and pigment structure in dermoscopy. The pathological bank is necessarily noisy because only image-level labels are available; the paper states that $y_x \in \{0,1\}$8 contains a mixture of normal and abnormal regions but remains biased toward the pathology distribution (Lian et al., 14 May 2026). This design choice is central to the framework’s later contrastive reasoning.

3. Evidence retrieval and evidence-aware reasoning

For a test image $y_x \in \{0,1\}$9, EviScreen first computes its patch feature map

$f : \mathbb{R}^{H \times W \times C} \to [0,1]$0

Each feature $f : \mathbb{R}^{H \times W \times C} \to [0,1]$1 is treated as a query patch (Lian et al., 14 May 2026). For each query patch, the method retrieves its $f : \mathbb{R}^{H \times W \times C} \to [0,1]$2 nearest neighbors from both knowledge banks:

$f : \mathbb{R}^{H \times W \times C} \to [0,1]$3

$f : \mathbb{R}^{H \times W \times C} \to [0,1]$4

Thus, $f : \mathbb{R}^{H \times W \times C} \to [0,1]$5 and $f : \mathbb{R}^{H \times W \times C} \to [0,1]$6 are evidence tensors in $f : \mathbb{R}^{H \times W \times C} \to [0,1]$7, representing normal and pathological historical evidence for every region of the current image (Lian et al., 14 May 2026).

The main predictive pathway is the evidence-aware reasoning module, implemented as a transformer-style architecture. The normal and pathological branches are initialized by copying the query patch features:

$f : \mathbb{R}^{H \times W \times C} \to [0,1]$8

For each layer, EviScreen performs cross-attention between the current patch representation and its retrieved evidence. In the normal branch,

$f : \mathbb{R}^{H \times W \times C} \to [0,1]$9

with an analogous expression for the pathological branch using $\hat{y}$0 (Lian et al., 14 May 2026). This aligns each query region with its nearest historical evidence and produces an evidence-aware patch representation.

After cross-attention, the model reshapes the patch grid into a sequence and applies inter-patch self-attention:

$\hat{y}$1

This allows the model to integrate spatial context across the entire image, such as multi-region disease patterns in chest radiographs or distributed retinal abnormalities (Lian et al., 14 May 2026).

After $\hat{y}$2 layers, the final normal and pathological representations are aggregated by [CLS] tokens, and prediction is produced by an MLP:

$\hat{y}$3

The reported reasoning module uses embedding dimension 1024, 8 heads, depth 4, and non-learnable 2D sin-cos positional encodings (Lian et al., 14 May 2026). The paper states that the module is trained for binary classification over $\hat{y}$4 and $\hat{y}$5; it further states that the training loss is standard binary cross-entropy, although the explicit formula is not written in the paper (Lian et al., 14 May 2026).

EviScreen also includes a training-free variant, EviScreen-TF, built from the same dual banks and frozen feature extractor but without the reasoning transformer. For each patch, it computes the average distance to the top-$\hat{y}$6 neighbors in the normal and pathological banks:

$\hat{y}$7

$\hat{y}$8

It then forms an abnormality map by contrast:

$\hat{y}$9

A high value indicates that a region is farther from normal patterns than from pathological ones (Lian et al., 14 May 2026).

4. Interpretability by evidence and localization

A central claim of EviScreen is that it replaces post-hoc interpretability with evidence-native interpretability (Lian et al., 14 May 2026). The paper distinguishes two forms.

The first is retrospection interpretability. For any query region in a test image, the system can show the top retrieved normal patches and the top retrieved pathological patches, together with their source images and clinical labels. The examples reported include a fundus image with a large optic cup, for which normal evidence consists of normal optic discs and pathological evidence consists of discs with increased cup-disc ratio and glaucoma-like appearance; a chest X-ray with Pleural Effusion, for which normal evidence shows clear lung bases and pathological evidence shows similar effusion patterns; and dermoscopic lesions, for which benign nevus patterns are contrasted with BCC or melanoma patterns (Lian et al., 14 May 2026). This makes the model’s decision path explicitly case-based: the explanation is not only that a region mattered, but that it resembled specific historical normal and pathological regions.

The second is localization interpretability. Instead of deriving heat maps by backpropagation, EviScreen uses the contrastive abnormality map from Equations (8a), (8b), and (9). The paper reports that these maps are qualitatively sharper and more focused than deviation maps from PatchCore* built on the same foundation-model features, with examples showing precise highlighting of optic-disc abnormalities, hemorrhagic fundus regions, CXR effusions and consolidations, and dermoscopic tumor regions (Lian et al., 14 May 2026).

A quantitative reader study on 20 fundus images with lesion masks reports the following Dice and IoU scores for localization (Lian et al., 14 May 2026):

Method	Dice	IoU
PatchCore*	$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$0	$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$1
CIPL	$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$2	$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$3
EviScreen	$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$4	$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$5

The interpretability scheme therefore combines local evidence retrieval and contrastive localization. In the terminology of the paper, the first addresses “why” through historical cases, while the second addresses “where” through retrieval-derived abnormality structure rather than saliency (Lian et al., 14 May 2026).

5. Benchmarks and empirical performance

The evaluation protocol emphasizes both standard discrimination metrics and clinically oriented operating characteristics (Lian et al., 14 May 2026). In addition to AUROC and AP, the paper defines Specificity at $$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$6 Recall as

$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$7

$$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$8

with $$X_N = X_N^R \cup X_N^E, \qquad X_P = X_P^R \cup X_P^E,$$9 in the main experiments. It also defines Clear Separation Rate:

$X_N^R$0

CSR measures the fraction of samples whose scores lie outside the overlap region of the positive and negative score distributions (Lian et al., 14 May 2026).

Across JSIEC, RIADD, CheXpert, and Derm12345, the paper reports that EviScreen is best on AUROC, AP, Spe@95%R, Spe@99%R, Spe@100%R, and CSR (Lian et al., 14 May 2026). Selected results are summarized below.

External test set	AUROC	Clinically oriented results
JSIEC	EviScreen 98.06 vs FM 95.84	Spe@99%R 91.62 vs 79.29; Spe@100%R 91.27 vs 78.39; CSR 88.95 vs 80.02
RIADD	EviScreen 91.32 vs best baseline 90.42	Spe@99%R 59.39 vs FM 55.73; CSR 54.38 vs FM 48.91
CheXpert	EviScreen 96.72 vs FM 95.60	Spe@99%R 74.26 vs FM 67.66; CSR 70.59 vs FM 57.41
Derm12345	EviScreen 97.43 vs FM 95.93	Spe@99%R 85.78 vs FM 73.57; CSR 77.92 vs FM 55.09

The paper explicitly notes that some AUROC improvements are modest, whereas the gains in Spe@X%R and CSR are large (Lian et al., 14 May 2026). That pattern is consequential for screening, because the stated goal is high specificity at clinically high recall rather than general ranking quality alone.

The training-free variant also improves over deviation-based retrieval baselines. Figure 1(a), as described in the paper, shows that EviScreen-TF consistently outperforms PatchCore and PatchCore* on AUROC, AP, Spe@95%R, Spe@99%R, Spe@100%R, and CSR, even though the latter uses the same foundation-model features (Lian et al., 14 May 2026). Figure 1(b) further shows that performance improves as more samples are included in the knowledge banks, especially on the specificity-at-high-recall metrics (Lian et al., 14 May 2026).

Ablation results support the necessity of both main components. On JSIEC and RIADD, removing evidence retrieval or removing evidence-aware reasoning degrades AUROC, Spe@95%R, Spe@99%R, Spe@100%R, and CSR, while the full model gives the strongest results (Lian et al., 14 May 2026). This indicates that retrieval alone and reasoning alone are insufficient; the reported performance depends on their combination.

6. Training setup, deployment characteristics, limitations, and future directions

The evidential reasoning module is trained on frozen foundation-model features rather than by end-to-end finetuning of the backbones (Lian et al., 14 May 2026). The reported training configuration uses AdamW with weight decay $X_N^R$1, $X_N^R$2, $X_N^R$3, warm-up from one of $X_N^R$4 selected via validation, and cosine decay. Batch size is 32 for CFP and dermoscopy and 8 for CXR, and no data augmentation is used “to avoid creating or losing abnormalities” (Lian et al., 14 May 2026).

The implementation relies on PyTorch 2.4.1 and FAISS 1.8.0, with experiments run on NVIDIA RTX 3090 hardware (Lian et al., 14 May 2026). The reported memory footprint for the knowledge banks is approximately 1,002 MB for ophthalmology with 256k vectors, 4,004 MB for radiology with 1,024k vectors, and 766 MB for dermatology with 196k vectors. Retrieval time in the largest radiology setting is reported as approximately 0.6 s per image (Lian et al., 14 May 2026). The code is publicly available at https://github.com/DopamineLcy/EviScreen (Lian et al., 14 May 2026).

The paper identifies several limitations. The evaluation spans ten public datasets across three domains, but primarily 2D images; volumetric and temporal modalities are not considered (Lian et al., 14 May 2026). Generalization is demonstrated on specific external sets, but broader multicenter validation remains necessary. The pathological bank is contaminated by normal patches because only image-level labels are available, although a stress test with 20% extra normal samples added to the pathological bank shows that performance remains robust, and additional denoising via SoftPatch-style methods yields similar performance (Lian et al., 14 May 2026). The method also presupposes access to strong modality-specific foundation models, and it incurs nontrivial retrieval cost and bank storage.

The future directions stated in the paper include extension to more modalities, 3D medical images such as CT and MRI, and finer-grained screening tasks including multi-stage grading and multi-label prediction (Lian et al., 14 May 2026). The authors also suggest broader clinical deployment questions, including integration with PACS and privacy-preserving knowledge banks. A plausible implication is that EviScreen is best understood not as a generic post-hoc explanation layer, but as a retrieval-centered screening architecture in which predictive performance and interpretability are co-produced by the same evidence mechanism.