Pith. sign in

REVIEW 3 major objections 2 minor 24 references

A continual-learning version of PatchCore adapts medical anomaly detection to shifting data while forgetting less than 1% of prior knowledge.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A continual-learning variant of PatchCore matches task-specific models on medical image anomaly detection with less than 1% forgetting.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible first application of continual learning to medical VAD, but the abstract alone can't support the sub-1% forgetting claim without memory and task details. the 3 major comments →

arxiv 2508.18013 v1 pith:EAIX764Q submitted 2025-08-25 cs.CV cs.AI

Towards Continual Visual Anomaly Detection in the Medical Domain

Manuel Barusco , Francesco Borsatti , Nicola Beda , Davide Dalle Pezze , Gian Antonio Susto This is my paper
classification cs.CV cs.AI
keywords continual learningvisual anomaly detectionmedical imagingPatchCoreCLcatastrophic forgettingBMADmodel adaptationincremental learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that visual anomaly detection in medical imaging can be made continual: a model can keep learning from new normal data without catastrophically forgetting earlier data. The authors adapt PatchCore, a standard anomaly detection method, into PatchCoreCL and test it on BMAD, a medical imaging dataset with pixel-level annotations. They report performance on par with models trained separately on each task, with a forgetting value below 1%. If correct, this makes it feasible to deploy anomaly detection systems that track evolving medical image distributions without full retraining.

Core claim

The paper establishes that a continual learning setup for medical visual anomaly detection is feasible. PatchCoreCL, a continual version of the established PatchCore model, when trained on a sequence of tasks drawn from the BMAD medical imaging dataset, achieves detection and localization performance comparable to task-specific models while exhibiting a forgetting value of less than 1%. This is, per the authors, the first demonstration of VAD in a continual learning scenario for the medical field, suggesting that incremental adaptation is a viable path for maintaining model accuracy as clinical imaging data evolves.

What carries the argument

PatchCoreCL, a continual-learning variant of the PatchCore anomaly-detection model. PatchCore works by storing normal feature representations in a memory bank and comparing new images against that memory; PatchCoreCL updates this memory incrementally across tasks, which is what allows new distributions to be learned without retraining from scratch and with minimal forgetting of previously seen normal patterns.

Load-bearing premise

The evaluation assumes that a fixed sequence of tasks built from the BMAD dataset faithfully represents the gradual, unlabeled distribution changes that occur in real medical imaging; if the task boundaries are artificial or the shifts are not representative, the sub-1% forgetting may not transfer to clinical settings.

What would settle it

Train PatchCoreCL on a sequence of medical imaging tasks where distribution shift is gradual and unlabeled (for example, images from a new scanner arriving without task labels) and measure forgetting; if forgetting exceeds a few percent or detection accuracy drops markedly, the claim of sub-1% forgetting in realistic conditions would be falsified. Alternatively, re-running the same BMAD task sequence with different task orderings and observing large variability in forgetting would suggest the reported figure is an artifact of the chosen split.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Medical VAD models can be updated incrementally as imaging protocols or patient populations shift, without full retraining.
  • Forgetting below 1% means previously learned anomaly types remain detectable after learning new ones, preserving diagnostic safety.
  • A single continually trained model could replace a suite of task-specific models, reducing storage and compute overhead in clinical deployment.
  • The same continual approach could be extended to other medical imaging datasets and modalities, provided task sequences reflect real-world drift.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported sub-1% forgetting likely depends on how the BMAD tasks are split and ordered; if real-world drift is gradual and unlabeled, the model may need additional mechanisms for detecting task boundaries to match these numbers.
  • Because PatchCore is largely memory-based and non-parametric, it may be inherently resistant to catastrophic forgetting; other VAD architectures, such as deep reconstruction networks, may not enjoy the same benefit in a continual setting.
  • A direct extension would be to test PatchCoreCL on a sequence of tasks with varying degrees of distribution shift (e.g., different scanners, contrast agents) and measure whether forgetting scales with task dissimilarity.
  • Since BMAD provides pixel-level annotations, a useful check is whether localization quality stays stable across tasks, not just image-level detection accuracy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The manuscript, based on the abstract supplied, proposes PatchCoreCL, a continual learning variant of the PatchCore anomaly detector, for medical visual anomaly detection. The authors evaluate PatchCoreCL on BMAD, a real-world medical imaging dataset, and report that it achieves performance 'comparable' to task-specific models with a 'forgetting value less than a 1%.' The abstract frames this as the first application of continual learning to medical visual anomaly detection.

Significance. If the central claim is substantiated, this work would provide a valuable first benchmark for continual visual anomaly detection in medical imaging, a domain where distribution shift is a practical concern. The claim is concrete and testable: comparable AUC/pixel-level metrics to task-specific models and a sub-1% forgetting value. However, as submitted, the manuscript contains only the abstract; none of the experimental details, definitions, or numbers needed to verify this claim are provided. The contribution cannot currently be assessed beyond its plausibility.

major comments (3)
  1. [Abstract] The central quantitative claim, 'forgetting value less than a 1%,' is stated without reporting the actual measured value, error bars, or number of runs. 'Comparable performance' similarly lacks any numeric comparison (e.g., image-level AUC and pixel-level AUROC for PatchCoreCL versus each task-specific baseline). Without these numbers, the headline claim is unsupported and cannot be checked.
  2. [Abstract] PatchCore is a memory-based detector that relies on a coreset of normal features. The continual variant PatchCoreCL must specify how the coreset is updated across tasks. If coresets are accumulated without a memory bound, sub-1% forgetting is expected from rehearsal and does not demonstrate continual stability under resource constraints. The manuscript should state whether the coreset is bounded, how old data are summarized or discarded, and report memory usage as a function of task count.
  3. [Abstract] The evaluation protocol for BMAD is undefined. The abstract does not describe how the dataset is split into tasks, the ordering of tasks, or the type of continual learning setting (task-incremental, domain-incremental, class-incremental). Forgetting values depend critically on task order and similarity. The authors must define the task sequence and report per-task metrics after each learning step, not merely an aggregate forgetting value.
minor comments (2)
  1. [Abstract] The forgetting metric should be defined formally (e.g., average per-task AUC drop after learning the next task). Standard continual learning metrics (backward transfer, forward transfer) should also be reported.
  2. [Abstract] The 'task-specific models' baseline should be clarified: does each task-specific model use the same PatchCore architecture and training set? Further, the claim of 'for the first time' needs a brief related-work comparison to substantiate novelty.

Circularity Check

0 steps flagged

No significant circularity detected from the available text

full rationale

The only evidence available is the abstract. The central claim—that PatchCoreCL achieves performance comparable to task-specific models with under 1% forgetting—is an empirical result evaluated on BMAD, an external public medical imaging dataset, and compared against PatchCore, a published baseline model. There is no fitted parameter renamed as a prediction, no equation reused as its own conclusion, and no load-bearing self-citation in the abstract. Concerns about the continual-learning protocol (e.g., memory bounds, task splitting, definition of forgetting) pertain to experimental validity and generalizability, not to circularity of the derivation chain. With no quoted reduction of the claimed result to its own inputs, the appropriate finding is no circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The central claim rests on the representativeness of the BMAD continual learning setup and the validity of PatchCore's normal-only memory under distribution shift. No free parameters or invented entities are visible from the abstract alone, but the full paper likely introduces task ordering and hyperparameter choices that would need to be audited.

axioms (3)
  • domain assumption BMAD's task structure is a valid proxy for real-world medical distribution shift.
    The paper evaluates PatchCoreCL on BMAD in a continual learning scenario, and the abstract gives no details about how tasks are split or ordered. If this split does not reflect real clinical drift, the forgetting result will not transfer.
  • domain assumption PatchCore's normal-only memory bank remains effective for medical anomaly detection under distribution shift.
    The method is a continual learning version of PatchCore, which detects anomalies by comparing test features against a memory bank of normal features. This assumption is imported from the original PatchCore work and is not examined in the abstract.
  • domain assumption The forgetting metric is computed using a standard continual learning protocol.
    The claim of 'forgetting value less than a 1%' depends on a specific definition of forgetting and a specific evaluation protocol. The abstract does not provide the metric definition, so this remains an unverified measurement assumption.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Continual Visual Anomaly Detection in the Medical Domain." pith.science (2026). https://pith.science/paper/EAIX764Q

@misc{pith2026250818013,
  author       = {Pith},
  title        = {Pith review of: Towards Continual Visual Anomaly Detection in the Medical Domain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EAIX764Q}},
  note         = {Machine review of arXiv:2508.18013}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Visual Anomaly Detection (VAD) seeks to identify abnormal images and precisely localize the corresponding anomalous regions, relying solely on normal data during training. This approach has proven essential in domains such as manufacturing and, more recently, in the medical field, where accurate and explainable detection is critical. Despite its importance, the impact of evolving input data distributions over time has received limited attention, even though such changes can significantly degrade model performance. In particular, given the dynamic and evolving nature of medical imaging data, Continual Learning (CL) provides a natural and effective framework to incrementally adapt models while preserving previously acquired knowledge. This study explores for the first time the application of VAD models in a CL scenario for the medical field. In this work, we utilize a CL version of the well-established PatchCore model, called PatchCoreCL, and evaluate its performance using BMAD, a real-world medical imaging dataset with both image-level and pixel-level annotations. Our results demonstrate that PatchCoreCL is an effective solution, achieving performance comparable to the task-specific models, with a forgetting value less than a 1%, highlighting the feasibility and potential of CL for adaptive VAD in medical imaging.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 23 canonical work pages · 2 internal anchors

  1. [1]

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

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 9584--9592, 2019

    work page 2019

  2. [2]

    K. K. Santhosh, D. P. Dogra, and P. P. Roy. Anomaly detection in road traffic using visual surveillance: A survey. ACM Comput. Surv. , 53(6), December 2020

    work page 2020

  3. [3]

    BMAD: Benchmarks for Medical Anomaly Detection

    Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. Bmad: Benchmarks for medical anomaly detection. arXiv preprint arXiv:2306.11876 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    Towards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Sch\"olkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 14318--14328, June 2022

    work page 2022

  5. [5]

    Unveiling the anomalies in an ever-changing world: A benchmark for pixel-level anomaly detection in continual learning

    Nikola Bugarin, Jovana Bugaric, Manuel Barusco, Davide Dalle Pezze, and Gian Antonio Susto. Unveiling the anomalies in an ever-changing world: A benchmark for pixel-level anomaly detection in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4065--4074, 2024

    work page 2024

  6. [6]

    Moviad: A modular library for visual anomaly detection, 2025

    Manuel Barusco, Francesco Borsatti, Arianna Stropeni, Davide Dalle Pezze, and Gian Antonio Susto. Moviad: A modular library for visual anomaly detection, 2025

    work page 2025

  7. [7]

    Draem -- a discriminatively trained reconstruction embedding for surface anomaly detection, 2021

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem -- a discriminatively trained reconstruction embedding for surface anomaly detection, 2021

    work page 2021

  8. [8]

    Reconstruction by inpainting for visual anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition , 112:107706, 2021

    work page 2021

  9. [9]

    Paste: Improving the efficiency of visual anomaly detection at the edge

    Manuel Barusco, Francesco Borsatti, Davide Dalle Pezze, Francesco Paissan, Elisabetta Farella, and Gian Antonio Susto. Paste: Improving the efficiency of visual anomaly detection at the edge. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops , pages 4026--4035, June 2025

    work page 2025

  10. [10]

    Memory efficient continual learning for edge-based visual anomaly detection, 2025

    Manuel Barusco, Lorenzo D'Antoni, Davide Dalle Pezze, Francesco Borsatti, and Gian Antonio Susto. Memory efficient continual learning for edge-based visual anomaly detection, 2025

    work page 2025

  11. [11]

    Experience replay for continual learning

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems , volume 32. Curran Associates, Inc., 2019

    work page 2019

  12. [12]

    Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences , 114(13):352...

    work page 2017

  13. [13]

    Learning without forgetting, 2017

    Zhizhong Li and Derek Hoiem. Learning without forgetting, 2017

    work page 2017

  14. [14]

    Rusu, Alexander Pritzel, and Daan Wierstra

    Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks, 2017

    work page 2017

  15. [15]

    Packnet: Adding multiple tasks to a single network by iterative pruning, 2018

    Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning, 2018

    work page 2018

  16. [16]

    Latent replay for real-time continual learning, 2020

    Lorenzo Pellegrini, Gabriele Graffieti, Vincenzo Lomonaco, and Davide Maltoni. Latent replay for real-time continual learning, 2020

    work page 2020

  17. [17]

    Learn to detect objects incrementally

    Linting Guan, Yan Wu, Junqiao Zhao, and Chen Ye. Learn to detect objects incrementally. In 2018 IEEE Intelligent Vehicles Symposium (IV) , pages 403--408. IEEE, 2018

    work page 2018

  18. [18]

    Towards continual adaptation in industrial anomaly detection

    Wujin Li, Jiawei Zhan, Jinbao Wang, Bizhong Xia, Bin-Bin Gao, Jun Liu, Chengjie Wang, and Feng Zheng. Towards continual adaptation in industrial anomaly detection. In Proceedings of the 30th ACM International Conference on Multimedia , pages 2871--2880, 2022

    work page 2022

  19. [19]

    Continual Learning Approaches for Anomaly Detection

    Davide Dalle Pezze, Eugenia Anello, Chiara Masiero, and Gian Antonio Susto. Continual learning approaches for anomaly detection. arXiv preprint arXiv:2212.11192 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    Student-teacher feature pyramid matching for anomaly detection

    Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student-teacher feature pyramid matching for anomaly detection. arXiv preprint arXiv:2103.04257 , 2021

    Pith/arXiv arXiv 2021

  21. [21]

    Efficientad: Accurate visual anomaly detection at millisecond-level latencies, 2024

    Kilian Batzner, Lars Heckler, and Rebecca König. Efficientad: Accurate visual anomaly detection at millisecond-level latencies, 2024

    work page 2024

  22. [22]

    Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows, 2021

    Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows, 2021

    work page 2021

  23. [23]

    Unsupervised continual anomaly detection with contrastively-learned prompt

    Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proceedings of the AAAI conference on artificial intelligence , volume 38, pages 3639--3647, 2024

    work page 2024

  24. [24]

    Wide residual networks, 2017

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks, 2017

    work page 2017

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.