Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Quantitative Benchmarking of Anomaly Detection Methods in Digital Pathology

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Benchmarking 23 anomaly detection methods on pathology images, this paper finds feature-distribution methods lead on average while scale and epoch selection shift the rankings.

desk verdict Useful pathology-specific benchmark, but the family ranking rests on uniform defaults and oracle validation; needs sensitivity checks and error bars. read the letter →

arxiv 2506.19234 v1 pith:7MOXCNAW submitted 2025-06-24 eess.IV cs.CV

Can Cui , Xindong Zheng , Ruining Deng , Quan Liu , Tianyuan Yao , Keith T Wilson , Lori A Coburn , Bennett A Landman
show 3 more authors
Haichun Yang Yaohong Wang Yuankai Huo
This is my paper
classification eess.IVcs.CV
keywords anomalydetectiondigitalpathologybenchmarkwholeslideimagesfeaturedistributionmethodsepochselectionimagescaleROC-AUC
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Anomaly detection in digital pathology has been borrowed from industrial defect inspection without a clear picture of which methods actually work on tissue images. This paper benchmarks 23 anomaly detection methods across five datasets, including real pathology, synthetic pathology, and industrial images, under a controlled protocol that varies image scale, anomaly pattern type, and training epoch selection strategy. It finds that no method wins everywhere, but on average feature-distribution methods outperform distillation-based and reconstruction-based methods on pathology images. It also finds that epoch selection matters a lot, with the ranking VAL > sample-wise > loss > last, and that patch scale changes results more than the multi-scale features built into many methods. These results are intended as a baseline for researchers choosing or designing anomaly detectors for pathology.

What carries the argument

The benchmark protocol is the argument-carrying mechanism. Twenty-three methods are grouped by algorithmic family and run under a standardized setup: 256x256 input patches, 200 training epochs, batch size 16, default hyperparameters, and ImageNet-standard normalization. Each method outputs a patch-level anomaly score, and object-level scores are aggregated as the average of the top 20% patch scores. Four epoch-selection strategies are compared: per-sample best checkpoint (sample-wise), best checkpoint on a complete validation set with abnormal samples (VAL), minimum loss on normal validation data (LOSS), and the last epoch (LAST). Three real pathology datasets, two synthetic pathology datasets with controlled circle/cross and density anomalies, and two industrial datasets provide the test bed, with ROC-AUC and PR-AUC as metrics. This controlled comparison is what lets the authors attribute performance differences to method family, image scale, anomaly pattern, and epoch-selection strategy.

What would settle it

Re-run the same 23 methods on the same five datasets with per-dataset hyperparameter tuning; if the average family ranking changes materially—for example, if distillation-based methods rise above feature-distribution methods, or if the VAL > sample-wise ordering flips—the reported rankings are not intrinsic to the methods but to the evaluation protocol.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is an empirical ranking: when 23 anomaly detection methods from five families—reconstruction, feature distribution, distillation, normalizing flow, and synthetic anomaly—are run on real and synthetic pathology datasets with a unified protocol, feature distribution methods achieve the highest average object-level ROC-AUC (0.692) with the complete validation set for epoch selection, versus 0.600 for distillation-based and 0.564 for reconstruction-based methods. The paper further reports that no single method dominates all datasets, that the ordering of epoch-selection strategies is consistently VAL > sample-wise > LOSS > LAST, that smaller scales help for local patterns while larger scales help for density patterns, and that reversing which pattern is called normal changes performance substantially. The authors present these results as a quantitative benchmark to guide anomaly detection development for pathology images, where large image sizes, multi-scale structures, stain variability, and repetitive normal patterns make generic methods behave differently than on industrial images.

Load-bearing premise

The load-bearing premise is that using one set of default hyperparameters for every method and every dataset is a fair comparison; if some methods are much more sensitive to hyperparameters than others, the reported family rankings could be an artifact of that uniform setting.

Editorial extensions

If this is right

  • For pathology images, feature-distribution methods are the safer default choice when no per-dataset tuning is available.
  • An epoch-selection strategy should be chosen deliberately, since it can change results as much as the choice of method; using even a partial validation set with abnormal samples beats all unsupervised strategies.
  • Patch scale should be matched to the anomaly type: local pattern anomalies favor small patches, density anomalies favor large patches, and multi-scale feature extraction inside a method does not fully compensate.
  • The asymmetric difficulty of normal versus abnormal patterns means anomaly detection results are not symmetric under relabeling, and reconstruction-based methods can reconstruct abnormal tissue while blurring normal tissue.
  • The benchmark provides a reusable baseline, but its average rankings carry the caveat of uniform default hyperparameters.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct consequence left implicit is that per-dataset hyperparameter tuning could reorder the family-level ranking; the reported averages describe default-configuration behavior, not method ceilings.
  • Because local patterns favor small patches and density anomalies favor large patches, a scale-ensemble that pools decisions from both scales is a natural next test that the paper's own data motivate but do not run.
  • The reversal experiment implies that when deploying in a new tissue type, the diversity of the class labeled normal may matter as much as the design of anomaly targets; measuring normal-class intra-class variance could predict which methods will struggle.
  • The sample-wise strategy's second-place showing suggests a cheaper hybrid—training-loss convergence for early stopping plus sample-wise ratio scoring for final selection—could approach the full-validation upper bound at lower memory cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper presents an empirical benchmark of 23 anomaly detection methods spanning five methodological families (reconstruction-based, feature distribution-based, distillation-based, normalizing flow-based, and synthetic anomaly-based) on a collection of real and synthetic digital pathology datasets, together with two industrial datasets. The study evaluates the influence of image scale, anomaly pattern type, reversal of normal/abnormal definitions, and four training-epoch selection strategies. The principal claims are that no single method dominates, feature-distribution methods outperform distillation- and reconstruction-based methods on average, the epoch selection ranking is VAL > SAMPLE-WISE > LOSS > LAST, and image scale significantly affects detection performance.

Significance. If the reported rankings are robust, this benchmark would serve as a useful reference for anomaly detection in digital pathology, particularly the analysis of unbiased epoch selection and scale effects. The paper's strengths include the breadth of methods and datasets, the use of publicly available implementations from established repositories, and the candid acknowledgment of the uniform-hyperparameter limitation. The main quantitative claims, however, rest on single-run comparisons with default hyperparameters, so the central family-level ranking and the epoch selection ordering are not yet fully supported by the evidence presented.

major comments (3)
  1. [§IV.B and §V.A.1] The family-level ranking reported in Section V.A.1 (distribution-based 0.692, distillation-based 0.600, reconstruction-based 0.564) is presented as a property of the method families, but all methods are run with default hyperparameters from their original implementations, as stated in the Limitations section. Because method families differ in hyperparameter sensitivity (for example, reconstruction methods depend on bottleneck size and loss weighting, while feature-distribution methods may be comparatively robust), the observed ordering could be an artifact of the uniform configuration. Please provide a sensitivity analysis, at least for a few representative methods per family, to show that the family ordering persists under reasonable hyperparameter variation.
  2. [§V.A.1 and Table III] The benchmark reports a single ROC-AUC value per method, dataset, and epoch selection strategy, with no repeated trials, confidence intervals, or significance tests. Given the modest gaps between family averages (differences of 0.092 and 0.036) and the small test sets for some datasets (for example, Glomerulus large-scale has 69 normal and 65 abnormal test patches; Hazelnut has 31 normal and 58 abnormal), these differences could be within sampling noise. Please report results over multiple seeds (at least three) or provide bootstrap confidence intervals, and state whether the family ordering and the ranking VAL > SAMPLE-WISE > LOSS > LAST remain stable.
  3. [§V.A.3] The epoch selection ranking is derived from the VAL strategy, which uses a complete validation set containing abnormal samples; this strategy is an oracle upper bound by design, so its superiority is unsurprising. The practically relevant unbiased comparison is SAMPLE-WISE versus LOSS, and the paper states that sample-wise performs only 'slightly better' without quantifying the difference. Without error bars or a significance test, the practical advantage of SAMPLE-WISE over LOSS is not established. Please clarify the practical significance of this distinction and quantify its uncertainty.
minor comments (5)
  1. [Abstract and Introduction] The abstract states 'over 20 classical and prevalent anomaly detection methods' while the Introduction states '23 anomaly detection methods'; please make the count consistent.
  2. [Table II and §V.A.3] Table II lists nine dataset rows, but Section V.A.3 refers to '20 datasets'; please clarify the dataset counting, for instance whether each pattern/scale combination is treated as a separate dataset, so the reader can map the reported averages to the table.
  3. [§IV.B] The 'average of the top 20% highest anomaly scores' aggregation rule is introduced without justification; since this is a free parameter that can affect results, please discuss its influence or cite a source that supports this choice.
  4. [§V.A.1] The term 'object-level ROC-AUC' is used in Section V.A.1 but is never explicitly defined; please define it in relation to patch-level and WSI-level scoring so that the reported averages are unambiguous.
  5. [§IV.B] The sentence 'Here is a refined version:' followed by a restatement of the ROC-AUC definition appears to be an editing leftover and should be removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark conclusions are empirical measurements on held-out data, not derivations that reduce to their inputs.

full rationale

This paper is an empirical benchmark rather than a derivation, so the circularity patterns that apply to claimed first-principles predictions mostly do not apply here. The central claims are measured ROC-AUC and PR-AUC values on held-out test data across five pathology datasets and two industrial datasets, using 23 externally published anomaly detection methods. The results are reported as observations (e.g., family-level averages of 0.692 vs. 0.600 vs. 0.564), and no equation in the paper defines a prediction in terms of its own output. The epoch-selection comparison is the closest thing to a methodological derivation, and it is not circular: Strategy 2 (VAL) is explicitly described as an oracle or upper bound because it uses a complete validation set containing abnormal labels, and the paper does not pretend this strategy is unbiased; it ranks first by construction because it selects the epoch with the highest validation ROC-AUC. Strategies 1 and 3 use only normal validation data and are evaluated against the test set, so their observed ranking is empirical. The sample-wise strategy is attributed to PANDA [20] and the four-strategy framework to the authors' prior work [47], but those citations supply the experimental protocol, not the conclusion; the paper's findings are produced by running the protocol on new data, not by assuming the cited results. The self-citation to [47] is therefore not load-bearing in the sense of forcing the reported ranking. The Limitations section candidly states that default parameters were used uniformly across datasets and that models were not fine-tuned per dataset; this is a legitimate threat to the fairness or generalizability of the family-level ranking, but it is not circularity, because the benchmark result is not identical to any fitted input or to the definition of the methods. No self-definitional equivalence, fitted-input-called-prediction, or imported-uniqueness step was identified.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities. It relies on hand-chosen experimental settings (aggregation percentile, patch size, default hyperparameters) and on assumptions about how representative the synthetic datasets are and how well validation sets match test sets. No parameters are fitted to data; the listed free parameters are design choices.

free parameters (3)
  • Top 20% anomaly score aggregation percentile = 20%
    Chosen by hand to convert patch-level scores to object/patient-level scores (Section IV.B). This percentile affects all reported AUC values and is not varied.
  • Input patch size = 256x256 pixels
    All models use 256x256 input patches (Section IV.B). This uniform setting interacts with the scale comparison and may favor some methods over others.
  • Training epochs and batch size = 200 epochs, batch size 16
    Fixed across all datasets and methods (Section IV.B). The choice of 200 epochs and batch size 16 is arbitrary and may disadvantage methods that need more or fewer epochs.
assumptions (3)
  • domain assumption Synthetic circle, cross, and density patterns are representative of pathology anomalies
    The synthetic datasets are used to study scale and reversion effects (Section IV.A.2). If these patterns do not realistically mimic tissue anomalies, the conclusions about pattern type and scale may not transfer to real pathology.
  • domain assumption Validation set distribution matches the test set
    The VAL epoch selection strategy assumes the validation set is representative of the test set (Section IV.B.2). For the Hazelnut and Tile datasets, the validation set is extracted from the test set, which violates independence.
  • domain assumption ROC-AUC is the most reliable metric for these test sets
    The authors choose ROC-AUC as the primary metric because the test sets are 'relatively balanced' (Section V.A.3). If the class imbalance is larger than assumed, conclusions based on ROC-AUC could differ from PR-AUC-based conclusions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantitative Benchmarking of Anomaly Detection Methods in Digital Pathology." pith.science (2026). https://pith.science/paper/7MOXCNAW

@misc{pith2026250619234,
  author       = {Pith},
  title        = {Pith review of: Quantitative Benchmarking of Anomaly Detection Methods in Digital Pathology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7MOXCNAW}},
  note         = {Machine review of arXiv:2506.19234}
}
read the original abstract

Anomaly detection has been widely studied in the context of industrial defect inspection, with numerous methods developed to tackle a range of challenges. In digital pathology, anomaly detection holds significant potential for applications such as rare disease identification, artifact detection, and biomarker discovery. However, the unique characteristics of pathology images, such as their large size, multi-scale structures, stain variability, and repetitive patterns, introduce new challenges that current anomaly detection algorithms struggle to address. In this quantitative study, we benchmark over 20 classical and prevalent anomaly detection methods through extensive experiments. We curated five digital pathology datasets, both real and synthetic, to systematically evaluate these approaches. Our experiments investigate the influence of image scale, anomaly pattern types, and training epoch selection strategies on detection performance. The results provide a detailed comparison of each method's strengths and limitations, establishing a comprehensive benchmark to guide future research in anomaly detection for digital pathology images.

Figures

Figures reproduced from arXiv: 2506.19234 by the authors.

Figure 1
Figure 1. Key characteristics and clinical applications of anomaly detection in digital pathological images. Compared to natural images, anomaly detec￾tion in pathological images has the following characteristics: 1) Greater scale variation in multiscale, including cells and tissues. 2) Diversity in staining, such as differences in staining intensity and imaging brightness. 3) Repetitive patterns, with blurry boundaries betwe… view at source ↗
Figure 2
Figure 2. Classification of anomaly detection models across five methodologi￾cal families: reconstruction-based, feature distribution-based, distillation￾based, normalizing flow-based, and synthetic anomaly-based methods. Each family represents a distinct algorithmic principle for distinguishing anomalies from normal data. Reconstruction-based methods rely on detecting pixel-level differences, feature distribution methods mod… view at source ↗
Figure 3
Figure 3. Overview of the datasets and example images spanning diverse digital pathology and industrial imaging modalities. The datasets include real-world pathology images (e.g., colon, glomerulus, Camelyon), synthetic pathology images with designed patterns, and standard industrial datasets (e.g., Hazelnut, Tile). This breadth ensures comprehensive evaluation across different anomaly types, image scales, and domain complexi… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 35 canonical work pages

  1. [1]

    Deep industrial image anomaly detection: A survey,

    J. Liu, G. Xie, J. Wang, S. Li, C. Wang, F. Zheng, and Y . Jin, “Deep industrial image anomaly detection: A survey,”Machine Intelligence Research, vol. 21, no. 1, pp. 104–135, 2024

    2024

  2. [2]

    Machine learning for anomaly detection: A systematic review,

    A. B. Nassif, M. A. Talib, Q. Nasir, and F. M. Dakalbab, “Machine learning for anomaly detection: A systematic review,”Ieee Access, vol. 9, pp. 78 658–78 700, 2021

    work page 2021

  3. [3]

    Using an anomaly detection approach for the segmentation of colorectal cancer tumors in whole slide images,

    Q. Gu, C. Meroueh, J. Levernier, T. Kroneman, T. Flotte, and S. Hart, “Using an anomaly detection approach for the segmentation of colorectal cancer tumors in whole slide images,”Journal of Pathology Informatics, vol. 14, p. 100336, 2023

    work page 2023

  4. [4]

    Learning image representations for anomaly detection: application to discovery of histological alterations in drug development,

    I. Zingman, B. Stierstorfer, C. Lempp, and F. Heinemann, “Learning image representations for anomaly detection: application to discovery of histological alterations in drug development,”Medical Image Analysis, vol. 92, p. 103067, 2024

    work page 2024

  5. [5]

    Automated anomaly detection in histology images using deep learning,

    L. Shelton, R. Soans, T. Shah, T. Forest, K. Janardhan, M. Napolitano, R. Gonzalez, G. Carlson, J. K. Shah, and A. Chen, “Automated anomaly detection in histology images using deep learning,” inMedical Imaging 2024: Digital and Computational Pathology, vol. 12933. SPIE, 2024, pp. 152–158

    work page 2024

  6. [6]

    A survey on unsupervised anomaly detection algorithms for industrial images,

    Y . Cui, Z. Liu, and S. Lian, “A survey on unsupervised anomaly detection algorithms for industrial images,”IEEE Access, vol. 11, pp. 55 297–55 315, 2023

    work page 2023

  7. [7]

    A survey on visual anomaly detection: Challenge, approach, and prospect,

    Y . Cao, X. Xu, J. Zhang, Y . Cheng, X. Huang, G. Pang, and W. Shen, “A survey on visual anomaly detection: Challenge, approach, and prospect,” arXiv preprint arXiv:2401.16402, 2024

    arXiv 2024

  8. [8]

    Medianomaly: A comparative study of anomaly detection in medical images,

    Y . Cai, W. Zhang, H. Chen, and K.-T. Cheng, “Medianomaly: A comparative study of anomaly detection in medical images,”Medical Image Analysis, p. 103500, 2025

    work page 2025

Show all 48 references
  1. [9]

    Unsupervised pathology detection: a deep dive into the state of the art,

    I. Lagogiannis, F. Meissen, G. Kaissis, and D. Rueckert, “Unsupervised pathology detection: a deep dive into the state of the art,”IEEE transactions on medical imaging, vol. 43, no. 1, pp. 241–252, 2023

    2023

  2. [10]

    Bmad: Benchmarks for medical anomaly detection,

    J. Bao, H. Sun, H. Deng, Y . He, Z. Zhang, and X. Li, “Bmad: Benchmarks for medical anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4042–4053

    2024

  3. [11]

    A unified model for multi-class anomaly detection,

    Z. You, L. Cui, Y . Shen, K. Yang, X. Lu, Y . Zheng, and X. Le, “A unified model for multi-class anomaly detection,”Advances in Neural Information Processing Systems, vol. 35, pp. 4571–4584, 2022

    2022

  4. [12]

    Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 8330–8339

    2021

  5. [13]

    Deep one-class classifi- cation via interpolated gaussian descriptor,

    Y . Chen, Y . Tian, G. Pang, and G. Carneiro, “Deep one-class classifi- cation via interpolated gaussian descriptor,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 1, 2022, pp. 383–392

    2022

  6. [14]

    Denoising autoencoders for unsupervised anomaly detection in brain mri,

    A. Kascenas, N. Pugeault, and A. Q. O’Neil, “Denoising autoencoders for unsupervised anomaly detection in brain mri,” inInternational Conference on Medical Imaging with Deep Learning. PMLR, 2022, pp. 653–664

    2022

  7. [15]

    Constrained unsupervised anomaly segmentation,

    J. Silva-Rodr ´ıguez, V . Naranjo, and J. Dolz, “Constrained unsupervised anomaly segmentation,”Medical Image Analysis, vol. 80, p. 102526, 2022

    2022

  8. [16]

    Ganomaly: Semi- supervised anomaly detection via adversarial training,

    S. Akcay, A. Atapour-Abarghouei, and T. P. Breckon, “Ganomaly: Semi- supervised anomaly detection via adversarial training,” inComputer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14. Sprin...

    2018

  9. [17]

    Unsupervised anomaly localization with structural feature-autoencoders,

    F. Meissen, J. Paetzold, G. Kaissis, and D. Rueckert, “Unsupervised anomaly localization with structural feature-autoencoders,” inInterna- tional MICCAI Brainlesion Workshop. Springer, 2022, pp. 14–24

    2022

  10. [18]

    Dfr: Deep feature reconstruction for unsupervised anomaly segmentation,

    J. Yang, Y . Shi, and Z. Qi, “Dfr: Deep feature reconstruction for unsupervised anomaly segmentation,”arXiv preprint arXiv:2012.07122, 2020

    2012 arXiv

  11. [19]

    Trans- former based models for unsupervised anomaly segmentation in brain mr images,

    A. Ghorbel, A. Aldahdooh, S. Albarqouni, and W. Hamidouche, “Trans- former based models for unsupervised anomaly segmentation in brain mr images,” inInternational MICCAI Brainlesion Workshop. Springer, 2022, pp. 25–44

    2022

  12. [20]

    Panda: Adapting pretrained features for anomaly detection and segmentation,

    T. Reiss, N. Cohen, L. Bergman, and Y . Hoshen, “Panda: Adapting pretrained features for anomaly detection and segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 2806–2814

    2021

  13. [21]

    Towards total recall in industrial anomaly detection,

    K. Roth, L. Pemula, J. Zepeda, B. Sch ¨olkopf, T. Brox, and P. Gehler, “Towards total recall in industrial anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 14 318–14 328

    2022

  14. [22]

    Cfa: Coupled-hypersphere-based fea- ture adaptation for target-oriented anomaly localization,

    S. Lee, S. Lee, and B. C. Song, “Cfa: Coupled-hypersphere-based fea- ture adaptation for target-oriented anomaly localization,”IEEE Access, vol. 10, pp. 78 446–78 454, 2022

    2022

  15. [23]

    DFKDE - Anomalib Documentation,

    “DFKDE - Anomalib Documentation,” https://anomalib.readthedocs.io/ en/latest/markdown/guides\\/reference/models/image/dfkde.html

  16. [24]

    Probabilistic mod- eling of deep features for out-of-distribution and adversarial detection,

    N. A. Ahuja, I. Ndiour, T. Kalyanpur, and O. Tickoo, “Probabilistic mod- eling of deep features for out-of-distribution and adversarial detection,” arXiv preprint arXiv:1909.11786, 2019

    1909 arXiv

  17. [25]

    Padim: a patch dis- tribution modeling framework for anomaly detection and localization,

    T. Defard, A. Setkov, A. Loesch, and R. Audigier, “Padim: a patch dis- tribution modeling framework for anomaly detection and localization,” inInternational conference on pattern recognition. Springer, 2021, pp. 475–489

    2021

  18. [26]

    Anomaly detection via reverse distillation from one-class embedding,

    H. Deng and X. Li, “Anomaly detection via reverse distillation from one-class embedding,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 9737–9746

    2022

  19. [27]

    Revisiting reverse distillation for anomaly detection,

    T. D. Tien, A. T. Nguyen, N. H. Tran, T. D. Huy, S. Duong, C. D. T. Nguyen, and S. Q. Truong, “Revisiting reverse distillation for anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 24 511–24 520

    2023

  20. [28]

    Student-teacher feature pyra- mid matching for anomaly detection,

    G. Wang, S. Han, E. Ding, and D. Huang, “Student-teacher feature pyra- mid matching for anomaly detection,”arXiv preprint arXiv:2103.04257, 2021

    2021 arXiv

  21. [29]

    Recontrast: Domain-specific anomaly detection via contrastive reconstruction,

    J. Guo, S. Lu, L. Jia, W. Zhang, and H. Li, “Recontrast: Domain-specific anomaly detection via contrastive reconstruction,”Advances in Neural Information Processing Systems, vol. 36, pp. 10 721–10 740, 2023

    2023

  22. [30]

    Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows,

    J. Yu, Y . Zheng, X. Wang, W. Li, Y . Wu, R. Zhao, and L. Wu, “Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows,”arXiv preprint arXiv:2111.07677, 2021

    2021 arXiv

  23. [31]

    Cflow-ad: Real-time unsu- pervised anomaly detection with localization via conditional normalizing flows,

    D. Gudovskiy, S. Ishizaka, and K. Kozuka, “Cflow-ad: Real-time unsu- pervised anomaly detection with localization via conditional normalizing flows,” inProceedings of the IEEE/CVF winter conference on applica- tions of computer vision, 2022, pp. 98–107

    2022

  24. [32]

    Fully convo- lutional cross-scale-flows for image-based defect detection,

    M. Rudolph, T. Wehrbein, B. Rosenhahn, and B. Wandt, “Fully convo- lutional cross-scale-flows for image-based defect detection,” inProceed- ings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 1088–1097

    2022

  25. [33]

    Cutpaste: Self-supervised learning for anomaly detection and localization,

    C.-L. Li, K. Sohn, J. Yoon, and T. Pfister, “Cutpaste: Self-supervised learning for anomaly detection and localization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9664–9674

    2021

  26. [34]

    Anomaly detection in medical imaging with deep perceptual autoen- coders,

    N. Shvetsova, B. Bakker, I. Fedulova, H. Schulz, and D. V . Dylov, “Anomaly detection in medical imaging with deep perceptual autoen- coders,”IEEE Access, vol. 9, pp. 118 571–118 583, 2021

    2021

  27. [35]

    Multiscale generative model using regularized skip-connections and perceptual loss for anomaly detection in toxicologic histopathology,

    P. Zehnder, J. Feng, R. N. Fuji, R. Sullivan, and F. Hu, “Multiscale generative model using regularized skip-connections and perceptual loss for anomaly detection in toxicologic histopathology,”Journal of Pathology Informatics, vol. 13, p. 100102, 2022

    2022

  28. [36]

    Unsupervised anomaly detection in digital pathology using gans,

    M. Pocevi ˇci¯ut˙e, G. Eilertsen, and C. Lundstr¨om, “Unsupervised anomaly detection in digital pathology using gans,” in2021 IEEE 18th Interna- tional Symposium on Biomedical Imaging (ISBI). IEEE, 2021, pp. 1878–1882

    2021

  29. [37]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” inComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11- 14, 2016, Proceedings, Part II 14. Springer, 2016, pp. 694–711

    2016

  30. [38]

    Ganomaly: Semi-supervised anomaly detection via adversarial training,

    S. Akcay, A. Atapour-Abarghouei, and T. P. Breckon, “Ganomaly: Semi-supervised anomaly detection via adversarial training,” inAsian conference on computer vision. Springer, 2018, pp. 622–637

    2018

  31. [39]

    Unsupervised anomaly detection on histopathology images using adversarial learning and simulated anomaly,

    Y .-C. Lai and W.-T. Chu, “Unsupervised anomaly detection on histopathology images using adversarial learning and simulated anomaly,” inAnnual Conference on Medical Image Understanding and Analysis. Springer, 2024, pp. 357–371

    2024

  32. [40]

    Diffusion models for out-of-distribution detection in digital pathology,

    J. Linmans, G. Raya, J. van der Laak, and G. Litjens, “Diffusion models for out-of-distribution detection in digital pathology,”Medical Image Analysis, vol. 93, p. 103088, 2024

    2024

  33. [41]

    1399 h&e-stained sentinel lymph node sections of breast cancer patients: the camelyon dataset,

    G. Litjens, P. Bandi, B. Ehteshami Bejnordi, O. Geessink, M. Balkenhol, P. Bult, A. Halilovic, M. Hermsen, R. van de Loo, R. V ogelset al., “1399 h&e-stained sentinel lymph node sections of breast cancer patients: the camelyon dataset,”GigaScience, vol. 7, no. 6, p. giy065, 2018

    2018

  34. [42]

    Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer,

    B. E. Bejnordi, M. Veta, P. J. Van Diest, B. Van Ginneken, N. Karsse- meijer, G. Litjens, J. A. Van Der Laak, M. Hermsen, Q. F. Manson, M. Balkenholet al., “Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer,” J...

    2017

  35. [43]

    A cross-platform informatics system for the gut cell atlas: integrating clinical, anatomical and histological data,

    S. Bao, S. Chiron, Y . Tang, C. N. Heiser, A. N. Southard-Smith, H. H. Lee, M. A. Ramirez, Y . Huo, M. K. Washington, E. A. Scovilleet al., “A cross-platform informatics system for the gut cell atlas: integrating clinical, anatomical and histological data,” inMedical Imaging 2...

    2021

  36. [44]

    Glo-in-one: holistic glomerular detection, segmentation, and lesion characterization with large-scale web image mining,

    T. Yao, Y . Lu, J. Long, A. Jha, Z. Zhu, Z. Asad, H. Yang, A. B. Fogo, and Y . Huo, “Glo-in-one: holistic glomerular detection, segmentation, and lesion characterization with large-scale web image mining,”Journal of Medical Imaging, vol. 9, no. 5, pp. 052 408–052 408, 2022

    2022

  37. [45]

    Unitopatho, a labeled histopathological dataset for colorectal polyps classification and adenoma dysplasia grading,

    C. A. Barbano, D. Perlo, E. Tartaglione, A. Fiandrotti, L. Bertero, P. Cas- soni, and M. Grangetto, “Unitopatho, a labeled histopathological dataset for colorectal polyps classification and adenoma dysplasia grading,” in2021 IEEE International Conference on Image Processing (I...

    2021

  38. [46]

    Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection,

    P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9592–9600

    2019

  39. [47]

    Feasibility of universal anomaly detection without knowing the abnormality in medical images,

    C. Cui, Y . Wang, S. Bao, Y . Tang, R. Deng, L. W. Remedios, Z. Asad, J. T. Roland, K. S. Lau, Q. Liuet al., “Feasibility of universal anomaly detection without knowing the abnormality in medical images,” in Workshop on Medical Image Learning with Limited and Noisy Data. Sprin...

    2023

  40. [48]

    Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly de- tection,

    S. Akc ¸ay, A. Atapour-Abarghouei, and T. P. Breckon, “Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly de- tection,” in2019 International Joint Conference on Neural Networks (IJCNN). IEEE, 2019, pp. 1–8

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.