Ano-NAViLa: Pathology Anomaly Detector

• Ano-NAViLa is a pathology anomaly detection framework that uses a frozen vision-language model combined with expert-curated normal and abnormal pathology term pools.
• It models dual semantic relationships between normal images and pathology knowledge to detect anomalies in one-class settings.
• The framework achieves high performance and interpretability through semantic deviation scoring and robust evaluation across diverse datasets.

Ano-NAViLa is a pathology anomaly detection framework whose full name is Normal and Abnormal pathology knowledge-augmented Vision-LLM for Anomaly detection in pathology images. It is designed for one-class anomaly detection in pathology images, especially whole-slide images (WSIs), under the realistic setting in which only normal training data are available while diseased patterns are rare, heterogeneous, and often underrepresented. The method combines a frozen pathology-specific vision-LLM (VLM), expert-curated normal and abnormal pathology term pools, and a lightweight trainable multilayer perceptron (MLP) to detect and localize anomalous tissue regions. Its defining idea is to model two semantic relationships simultaneously—normal image ↔ normal pathology knowledge and normal image ↔ abnormal pathology knowledge—and then score test samples by deviation from the normal-reference patterns of both relations (Song et al., 21 Aug 2025).

1. Problem formulation and conceptual rationale

Ano-NAViLa addresses anomaly detection in computational pathology under four difficulties emphasized in the paper: a computational bottleneck associated with gigapixel WSIs and heavy anomaly-detection backbones, tissue diversity in which even normal tissue is structurally heterogeneous, domain shift across institutions, scanners, magnification, and staining protocols, and limited interpretability in black-box detectors (Song et al., 21 Aug 2025).

The framework departs from purely appearance-based outlier detection. Instead of treating anomalies only as generic deviations from normal image statistics, it uses data-driven knowledge from a pathology VLM pretrained on large pathology image-text corpora together with expert knowledge in the form of curated pathology terms. This suggests a semantic reformulation of one-class pathology anomaly detection: a sample is not judged only by whether its image embedding is atypical, but also by whether its image-conditioned relation to pathology concepts departs from the relations learned from normal tissue.

A central conceptual point is that Ano-NAViLa uses abnormal pathology knowledge without abnormal image supervision. Training uses only normal images, yet abnormal pathology terms are still injected through the abnormal term pool. The paper argues that the interaction between normal images and abnormal pathology terms acts as a kind of regularization, improving the representation of normality. A plausible implication is that the abnormal term pool functions less as a surrogate label space than as a semantic contrastive axis defined entirely within normal-data training.

2. Architecture and pathology knowledge representation

Ano-NAViLa consists of four components: a normal term pool, an abnormal term pool, a pretrained VLM with visual and text encoders, and a trainable MLP and anomaly scoring scheme (Song et al., 21 Aug 2025).

The two expert-curated term pools are

$T^N=\{t^N_i\}^{n_N}_{i=1}$

and

$T^A=\{t^A_j\}^{n_A}_{j=1},$

where $t^N_i$ and $t^A_j$ are the $i$-th normal and $j$-th abnormal pathology terms. For lymph node metastasis detection, the final pools contain 92 normal pathology terms and 48 abnormal pathology terms. The paper states that the normal terms describe normal tissue organization and structural variation, whereas abnormal terms describe structural and morphological alterations associated with disease progression. Prompt construction for the text encoder uses the explicit template

$\text{``an image showing KEYWORD''}.$

The base VLM is denoted by $\Gamma$, with visual encoder $\Gamma^V$ and text encoder $\Gamma^T$. The specific backbone is CONCH, used as a pathology VLM pretrained on large-scale pathology image-caption pairs and kept frozen throughout training and inference. For an input image $T^A=\{t^A_j\}^{n_A}_{j=1},$0, the image embedding is

$T^A=\{t^A_j\}^{n_A}_{j=1},$1

The normal and abnormal text embeddings are

$T^A=\{t^A_j\}^{n_A}_{j=1},$2

with each text embedding also in $T^A=\{t^A_j\}^{n_A}_{j=1},$3.

Image-text association is computed through cosine similarity $T^A=\{t^A_j\}^{n_A}_{j=1},$4. For the two pools separately, Ano-NAViLa defines softmax-normalized weights

$T^A=\{t^A_j\}^{n_A}_{j=1},$5

Because these weights are usually small, often within $T^A=\{t^A_j\}^{n_A}_{j=1},$6, the method amplifies them with an exponential factor before fusion. Each text embedding is scaled, concatenated with the image embedding, and processed by the trainable MLP $T^A=\{t^A_j\}^{n_A}_{j=1},$7:

$T^A=\{t^A_j\}^{n_A}_{j=1},$8

Each concatenated vector lies in $T^A=\{t^A_j\}^{n_A}_{j=1},$9. Averaging within each pool yields two representative text-augmented image embeddings, $t^N_i$0 and $t^N_i$1, summarizing the image’s relation to the normal and abnormal term pools, respectively.

The MLP is deliberately small: it has three linear layers, with ReLU activations in the first two, mapping

$t^N_i$2

Since CONCH is frozen, the MLP is the only trainable component. The paper interprets its role as structuring a latent space in which embeddings of normal image + normal terms cluster together, embeddings of normal image + abnormal terms also cluster together, and the two clusters remain separated.

3. Training objective, semantic structure, and anomaly scoring

Training uses only normal images, thereby preserving the anomaly-detection setting (Song et al., 21 Aug 2025). For a batch of $t^N_i$3 normal images, the model produces $t^N_i$4 and $t^N_i$5 for each image $t^N_i$6. The loss is

$t^N_i$7

with

$t^N_i$8

$t^N_i$9

$t^A_j$0

where $t^A_j$1 and $t^A_j$2.

The optimization objective encourages high similarity among all $t^A_j$3 embeddings, high similarity among all $t^A_j$4 embeddings, and low similarity between $t^A_j$5 and $t^A_j$6. In effect, Ano-NAViLa learns two stable normal-reference clusters from only normal data: one corresponding to the relation between normal images and normal pathology knowledge, and one corresponding to the relation between normal images and abnormal pathology knowledge.

At inference time, centroid embeddings are computed from normal validation images:

• $t^A_j$7, the centroid of normal text-augmented image embeddings;
• $t^A_j$8, the centroid of abnormal text-augmented image embeddings.

For a test image, Ano-NAViLa defines two deviation scores:

$t^A_j$9

and

$i$0

The patch-level anomaly score is then

$i$1

This dual-deviation construction is the method’s main scoring mechanism. The paper’s intuition is explicit: a normal image should remain close both to the normal-term centroid and to the centroid describing how normal images relate to abnormal terms. An abnormal image will tend to differ from both patterns, causing both deviations to rise. This suggests that abnormal terminology is used not as supervised anomaly evidence, but as a second normal-reference manifold learned from normal tissue alone.

For WSI-level analysis, Ano-NAViLa first scores non-overlapping patches, maps those scores back to patch locations to form a heatmap, and applies a $i$2 erosion operation for smoothing. Two WSI-level aggregation rules are used:

$i$3

and

$i$4

where $i$5 is the set of top 1% highest-scoring patches.

4. Datasets, implementation, and evaluation protocol

Ano-NAViLa was trained on GastricLN and externally tested on Camelyon16 (Song et al., 21 Aug 2025). GastricLN is a private gastric lymph node dataset with 808 WSIs, comprising 751 normal and 57 metastasis cases, digitized at 20x magnification and collected from two hospitals. Its split is:

• Training: 643 normal WSIs
• Validation: 50 normal WSIs
• Test: 58 normal + 57 metastasis WSIs

Camelyon16 is a public breast lymph node metastasis dataset used only for external testing, with 129 WSIs, including 80 normal and 49 metastasis cases, scanned at 40x magnification. The paper also defines Camelyon16$i$6 as 22 macro-metastasis WSIs ($i$7 mm tumor diameter) plus 80 normal WSIs. The cross-organ design is a central feature of the evaluation, since training occurs on gastric lymph nodes and external testing on breast lymph nodes from different institutions and acquisition settings.

Patch extraction uses different scales in the two cohorts:

• GastricLN: $i$8 patches
• Camelyon16: $i$9 patches

Implementation details are unusually compact. The prompt is "an image showing KEYWORD", the VLM backbone is CONCH, the optimizer is Adam, the learning rate is 0.001, batch size is 100, gradient accumulation updates occur after processing 100 batches, and training runs for 1 epoch. Centroids are computed from the validation set of GastricLN. Only the MLP is trainable.

The baseline set spans reconstruction-based, representation-based, student-teacher, diffusion-based, and VLM-based anomaly detection methods:

• GANomaly
• STFPM
• FastFlow
• CFA
• AnoDDPM
• EfficientAD
• AnomalyCLIP

Evaluation metrics are AUROC at patch level and AUROC and AUPR at WSI level. The paper reports 95% confidence intervals computed using 2000-fold bootstrapping with 100% resampling with replacement.

5. Empirical performance, interpretability, and ablation findings

On GastricLN, Ano-NAViLa achieved patch-level AUROC = 0.9681, outperforming all baselines. Selected comparator values reported in the paper are STFPM: 0.9538, FastFlow: 0.9242, CFA: 0.8881, AnoDDPM: 0.8860, EfficientAD: 0.8432, AnomalyCLIP: 0.3314, and GANomaly: 0.4182 (Song et al., 21 Aug 2025).

At the WSI level on GastricLN, using maximum score $j$0, Ano-NAViLa achieved AUROC = 0.9967 and AUPR = 0.9971. Using top-1% average $j$1, it achieved AUROC = 0.9894 and AUPR = 0.9904. The strongest baseline reported was STFPM, with AUROC 0.9779 / 0.9797 and AUPR 0.9728 / 0.9714, depending on the aggregation rule.

External evaluation on Camelyon16 is the paper’s main robustness test. Using $j$2, Ano-NAViLa obtained AUROC = 0.8594 and AUPR = 0.8309; using $j$3, it obtained AUROC = 0.7702 and AUPR = 0.7941. The paper states that STFPM achieved AUROC 0.7158 / 0.7324 and AUPR 0.6862 / 0.7038 on the same benchmark. On Camelyon16$j$4, the method was especially strong: with $j$5, AUROC = 0.9858 and AUPR = 0.9547; with $j$6, AUROC = 0.9761 and AUPR = 0.9699.

The paper repeatedly links these gains to the combined use of normal and abnormal pathology knowledge. In ablations over embedding composition, four variants were tested: text embeddings only $j$7, image embedding only $j$8, image plus normal text $j$9, and image plus normal plus abnormal text $\text{``an image showing KEYWORD''}.$0. The reported finding is that removing any component causes a drop in performance in nearly all settings, that $\text{``an image showing KEYWORD''}.$1 has the largest individual impact, that adding normal text helps especially on Camelyon16, and that adding abnormal text further improves performance on both datasets. The authors explicitly conclude that there is a synergistic effect between normal and abnormal text embeddings.

Interpretability is one of Ano-NAViLa’s central claims. Because scoring depends on similarity to explicit pathology terms, the model can identify which terms best match a high-anomaly patch or WSI. The paper reports that representative high-scoring patches from Camelyon16, along with their top-matching pathology terms, were reviewed by an experienced pathologist and found to be highly relevant to the image content and class labels. It further reports consistent similarity patterns across GastricLN and Camelyon16: within the normal pool, terms such as “Helper T lymphocyte” and “Small dormant lymphocytes” had higher similarity for normal patches in both datasets, while within the abnormal pool, “Tumor buds that emerge from medium-sized tubules” showed consistently high similarity specifically for abnormal samples at both patch and WSI levels.

The framework is also computationally lightweight at the trainable level. With batch size 28, the reported efficiency figures are 0.69M trainable parameters, 3.61 ms latency, and 2098 MiB peak GPU memory. The paper notes that this is the smallest number of trainable parameters among all compared methods and the second-best latency, with moderate memory usage due mainly to the VLM backbone.

6. Scope, limitations, and naming disambiguation

Ano-NAViLa is specifically a pathology anomaly detection system, not a navigation model. This distinction matters because similarly named works occupy very different research areas. NaVILA is a Legged Robot Vision-Language-Action Model for Navigation (Cheng et al., 2024), and NavA$\text{``an image showing KEYWORD''}.$2 is a hierarchical embodied navigation framework for long-horizon instruction-driven navigation (Zhang et al., 6 Aug 2025). The pathology paper defines Ano-NAViLa independently as a knowledge-augmented VLM for anomaly detection in pathology images (Song et al., 21 Aug 2025).

Several limitations are stated explicitly. First, construction of the pathology term pools is manual: the current normal and abnormal term sets are curated from the pathology literature and validated by an experienced pathologist. Second, broader external validation across more organs and datasets remains future work; the present experiments focus on lymph node metastasis detection across gastric and breast settings. Third, although the trainable component is small, the frozen VLM still contributes to memory cost, so improving computational efficiency—especially the VLM—remains an open direction.

Two further caveats concern interpretation. Ano-NAViLa is not simply “zero-shot CONCH with prompts”; the paper reports that zero-shot CONCH with the same term pools performed much worse than the proposed method, indicating that performance depends on the text-augmented representation and the training framework rather than on the backbone alone. Conversely, the method is not merely a generic visual outlier detector with post hoc explanations. Its anomaly score is constructed directly from deviations in two learned semantic relation spaces, and the text associations are part of the model’s forward computation rather than an auxiliary explanation layer.

In that sense, Ano-NAViLa is best understood as a semantically guided one-class anomaly detector for pathology. Its novelty lies in coupling a frozen pathology VLM with expert-curated normal and abnormal knowledge pools, using image-conditioned weighting to produce text-augmented image embeddings, learning two normal-reference semantic clusters from normal images only, and scoring anomalies by deviation from both clusters. The paper’s results indicate that this formulation improves anomaly detection, localization, robustness under domain shift, and pathology-grounded interpretability within the evaluated lymph node metastasis setting (Song et al., 21 Aug 2025).