acceptodds
Withdrawn from ICLR 2027

HiNorm: Hierarchical Normality Learning for Whole-Slide Image Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised anomaly detection (UAD) in whole-slide images (WSIs) provides a scalable approach to detecting pathological abnormalities by learning solely from normal WSIs, without requiring abnormal samples for training or lesion-level annotations. Existing methods, however, typically learn normality from isolated patches and derive slide-level predictions by aggregating patch anomaly scores. This patch-centric paradigm overlooks the spatial organization of tissue, even though pathological assessment depends on both local morphology and global architecture, and may therefore yield unreliable WSI-level decisions. We propose HiNorm, a hierarchical normality learning framework trained exclusively on normal WSIs. HiNorm employs dual-branch reverse distillation, with frozen patch-level and WSI-level pathology foundation models providing complementary references for local tissue morphology and whole-slide organization. Its key component, Asymmetric Cross-Level Reconstruction, uses global slide context to guide patch-level reconstruction while allowing fine-grained patch evidence to inform WSI-level reconstruction. This asymmetric interaction integrates complementary information without conflating the distinct pathological roles of the two representation levels. At inference, patch-level reconstruction discrepancies are mapped spatially for anomaly localization, whereas the WSI-level discrepancy directly measures slide-level abnormality, eliminating heuristic aggregation of patch scores. Across three public WSI benchmarks, HiNorm achieves state-of-the-art performance in both anomaly localization and WSI-level detection, while substantially reducing the WSI–patch performance gap.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Computational pathology, unsupervised anomaly detection, whole-slide images

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JKowOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HiNorm, a hierarchical normality learning framework for unsupervised anomaly detection in whole-slide images (WSIs). The method uses pretrained patch-level and WSI-level pathology foundation models to extract complementary representations. A dual-branch reverse distillation framework with asymmetric cross-level reconstruction is introduced to learn normal tissue patterns at both levels. Anomaly scores are computed from reconstruction discrepancies for patch-level localization and WSI-level detection.

Critical strengths

The proposed framework combines local organizational morphology with global WSI representation, utilizing a unified reconstruction framework and pre-trained pathological foundational models to provide a practically valuable solution for unsupervised anomaly detection.

Critical weaknesses
  1. The paper argues that existing patch-based approaches overlook spatial organization and global tissue structure because they aggregate patch-level anomaly scores for WSI-level prediction. However, graph-based methods and Transformer-based WSI models can capture spatial relationships and global context during representation learning, even when their final predictions rely on patch-level aggregation.

  2. In the patch reconstruction branch, multiple patch tokens attend to a single WSI token. Under standard softmax attention, a single key/value token produces attention weights that are always one, effectively broadcasting the same WSI information to all patch tokens. This raises questions about the actual role of the proposed cross-attention mechanism.

  3. TITAN already provides global tissue representations learned from large-scale pathology data. It is therefore unclear how much of the WSI-level improvement comes from the proposed normality learning framework rather than the strong pretrained representations.

  4. The framework mainly combines existing techniques, including reverse distillation, pretrained foundation models, dual-branch reconstruction, and cross-attention. Although their integration is effective, the proposed asymmetric reconstruction does not introduce a clearly distinct learning objective or explicitly constrain the relationship between patch-level and WSI-level normality.

Questions for authors

See weakness. In particular, clarifications regarding the single-token cross-attention mechanism and comparisons against anomaly detectors using frozen TITAN representations would be helpful.

Reviewer_zQwaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiNorm, an unsupervised anomaly‑detection method leveraging multi‑level foundation‑model features. It distinguishes patch‑level localization and WSI‑level detection tasks. Its core module, Asymmetric Cross‑Level Reconstruction (ACLR), employs bidirectional cross‑attention: global information participates in patch reconstruction while local information contributes to WSI reconstruction. This represents a promising research direction. The paper achieves strong experimental results, and performs especially well on WSI‑level anomaly detection and anomalous‑region localization, demonstrating the application potential of the proposed method.

Critical strengths

This paper proposes HiNorm, an unsupervised anomaly‑detection method leveraging multi‑level foundation‑model features. It distinguishes patch‑level localization and WSI‑level detection tasks. Its core module, Asymmetric Cross‑Level Reconstruction (ACLR), employs bidirectional cross‑attention: global information participates in patch reconstruction while local information contributes to WSI reconstruction. This represents a promising research direction. The paper achieves strong experimental results, and performs especially well on WSI‑level anomaly detection and anomalous‑region localization, demonstrating the application potential of the proposed method.

Critical weaknesses
  1. The novelty of the method and the necessity of the cross‑level interaction design require further justification. The paper highlights joint learning of patch‑level and WSI‑level normality as a key contribution. However, prior work has already investigated collaborative modeling of local and global normality. For instance, Yao et al. (2024, Engineering Applications of Artificial Intelligence) presented Local–global normality learning and discrepancy normalizing flow for unsupervised image anomaly detection. That work uses a two‑branch Transformer to reconstruct local and global normal features separately and introduces 40 global prototype tokens to aggregate global semantic information. Different from that work, HiNorm’s patch branch uses only one single WSI‑level token as the Key and Value for cross‑attention.

  2. In the ablation study shown in Table 3, the WSI I‑AUROC of the “WB Only” variant is 0.9641, which is even higher than 0.9581 obtained by full HiNorm. Therefore, further clarification is needed: does HiNorm’s performance advantage for WSI‑level detection mainly stem from explicit WSI normality learning, or from the cross‑level information interaction brought by ACLR? How substantial is the contribution of patch‑level information to WSI reconstruction? It would be helpful to construct a simple baseline that directly uses off‑the‑shell WSI‑level foundation‑model features combined with classic approaches such as K‑NN or Mahalanobis distance, and check whether such simple baselines can also achieve competitive performance. These comparisons can disentangle whether performance gains come from powerful pre‑trained WSI features, or from the specific network design of HiNorm.

  3. When anomalous tissue occupies only a very small fraction within a WSI, if most normal patches yield low reconstruction errors, could the averaging strategy dilute the responses originating from the few anomalous regions and thus degrade sensitivity for tiny lesions? The authors are suggested to conduct grouped evaluation according to the fraction of anomalous regions inside each full WSI.

  4. The paper has compared patch‑level feature extractors including DINOv2, Phikon, Hibou‑B and CONCH. Nevertheless, representative pathology foundation models such as UNI2‑h and Virchow2 are missing from comparisons. Moreover, TITAN is always adopted as the WSI‑level feature extractor. It remains unvalidated whether HiNorm can generalize to other WSI‑level foundation models.

Questions for authors
  1. The novelty of the method and the necessity of the cross‑level interaction design require further justification. The paper highlights joint learning of patch‑level and WSI‑level normality as a key contribution. However, prior work has already investigated collaborative modeling of local and global normality. For instance, Yao et al. (2024, Engineering Applications of Artificial Intelligence) presented Local–global normality learning and discrepancy normalizing flow for unsupervised image anomaly detection. That work uses a two‑branch Transformer to reconstruct local and global normal features separately and introduces 40 global prototype tokens to aggregate global semantic information. Different from that work, HiNorm’s patch branch uses only one single WSI‑level token as the Key and Value for cross‑attention.

  2. In the ablation study shown in Table 3, the WSI I‑AUROC of the “WB Only” variant is 0.9641, which is even higher than 0.9581 obtained by full HiNorm. Therefore, further clarification is needed: does HiNorm’s performance advantage for WSI‑level detection mainly stem from explicit WSI normality learning, or from the cross‑level information interaction brought by ACLR? How substantial is the contribution of patch‑level information to WSI reconstruction? It would be helpful to construct a simple baseline that directly uses off‑the‑shell WSI‑level foundation‑model features combined with classic approaches such as K‑NN or Mahalanobis distance, and check whether such simple baselines can also achieve competitive performance. These comparisons can disentangle whether performance gains come from powerful pre‑trained WSI features, or from the specific network design of HiNorm.

  3. When anomalous tissue occupies only a very small fraction within a WSI, if most normal patches yield low reconstruction errors, could the averaging strategy dilute the responses originating from the few anomalous regions and thus degrade sensitivity for tiny lesions? The authors are suggested to conduct grouped evaluation according to the fraction of anomalous regions inside each full WSI.

  4. The paper has compared patch‑level feature extractors including DINOv2, Phikon, Hibou‑B and CONCH. Nevertheless, representative pathology foundation models such as UNI2‑h and Virchow2 are missing from comparisons. Moreover, TITAN is always adopted as the WSI‑level feature extractor. It remains unvalidated whether HiNorm can generalize to other WSI‑level foundation models.

Reviewer_BVrsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HiNorm for unsupervised whole-slide image anomaly detection and localization. It models patch-level and WSI-level normality through dual-branch feature reconstruction with asymmetric cross-level interaction. Reconstruction discrepancies are then used for local anomaly localization and slide-level anomaly detection.

Critical strengths
  1. The paper includes informative figures that help illustrate the proposed framework and qualitative results.

  2. Experiments cover multiple datasets, various baselines, and both WSI-level and patch-level evaluations.

Critical weaknesses
  1. The paper presents the lack of WSI-level normality modeling as a major research gap. However, global anomaly modeling and local–global normality learning have already been explored in prior studies. The paper fails to identify a genuinely unresolved problem beyond these established ideas.

  2. The proposed framework largely combines existing techniques, including local–global feature modeling, dual-branch reconstruction, and cross-attention. The specific methodological contribution beyond these established designs remains unclear.

  3. The WSI-only branch achieves higher AUROC and AP than the full model, raising questions about the necessity of the proposed cross-level interaction. Moreover, the weak cross-center WSI detection performance (AUROC: 0.5642) does not convincingly support the claimed generalization capability.

Questions for authors

Please see the weaknesses above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.