Pith. sign in

REVIEW 3 major objections 6 minor 39 references

This paper claims that reconstruction shifts generated by masked latent diffusion on normal images can be learned as an image-level anomaly score, with residuals from the same process localizing anomalies.

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 ↗

T0 review · deepseek-v4-flash

2026-08-05 00:57 UTC pith:74SX6J2R

load-bearing objection DMD is a solid incremental step in medical anomaly detection; the reconstruction-shift classifier is new, but the paper's only domain-shift test omits the image-level score and no code is released. the 3 major comments →

arxiv 2608.00444 v1 pith:74SX6J2R submitted 2026-08-01 cs.CV

Reconstruction-Shift Discrimination via Mask-Guided Latent Diffusion for Medical Anomaly Detection

Yibo Wan , Jinyu Cai , Yunhe Zhang , Yi Bin , See-kiong Ng This is my paper
classification cs.CV
keywords unsupervised medical anomaly detectionlatent diffusionmask-guided reconstructionreconstruction-shift discriminationvector-quantized autoencoderimage-level anomaly scoreanomaly localization
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 paper tries to establish that the usual residual between a medical image and its reconstruction is too ambiguous to serve as a reliable whole-image anomaly score: expressive models can reproduce pathology, and benign variation can create false residuals. It proposes DMD, which first maps normal training images into a compact quantized latent space, then uses localized masks plus a latent diffusion model to create reconstructed counterparts of normal images. A classifier learns to tell original normal images from these reconstructed counterparts, giving a learned image-level anomaly score, while the same restoration process yields a pixel-level residual map. On five public datasets across brain MRI, breast ultrasound, and chest X-ray, DMD reports the best overall image- and pixel-level performance among the compared methods. If correct, this gives a normal-only training recipe for medical anomaly detection that does not rely on residual magnitude alone.

Core claim

DMD's central claim is that reconstruction shift—the difference between a normal image and its masked diffusion reconstruction—is a learnable, normal-only supervisory signal for image-level anomaly detection. The paper pairs each normal image with a reconstructed counterpart produced by masking random regions of a quantized latent code and denoising them, then trains a binary classifier on the pair. At test time, the classifier probability is the image-level score; the average residual over overlapping masks is the pixel-level map. Reported results include image-level AUC of 87.2%, 93.8%, and 84.3% on BraTS2021, BUSI, and VinDr-CXR, and pixel-level APpix of 78.9%, 49.5%, 56.4%, and 71.3% on

What carries the argument

The mechanism is two-stage. A vector-quantized autoencoder trained only on normal images gives a discrete latent grid; then random square masks select latent positions, Gaussian noise is added only there, a latent diffusion U-Net denoises the masked positions using unmasked positions as clean context, and the fixed decoder maps the cleaned latent back to an image. This reconstructed counterpart is paired with the original as a self-supervised classification task. The same masked-restoration path is reused at inference with overlapping grid masks, so one pipeline gives both a classifier-based image-level score and an aggregated residual map.

Load-bearing premise

The image-level score is only as good as the assumption that the reconstruction shifts produced by the fixed masking policy on normal images mirror the shifts real pathologies would cause; if real abnormal images are faithfully reconstructed or shift differently, the learned score will mislead.

What would settle it

Train DMD on normal images from one pathology population, such as BraTS FLAIR, and evaluate image-level AUC on a different anomaly type, such as stroke lesions from ATLAS. If the classifier score no longer separates normal from abnormal while the residual map still localizes lesions, the learned reconstruction-shift signal is tied to the training masking policy rather than to pathology. A second check: add normal images with unusual scanners or acquisition noise and see whether the classifier flags them as anomalous more often than a residual-aggregation baseline does.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Image-level anomaly scores can be learned directly from normal-only data, eliminating the need to hand-collapse residual maps into a scalar.
  • The same masked-restoration pass produces both outputs, so detection and localization can be trained jointly instead of as two separate models.
  • Because classification gradients flow into the denoiser, improving the discriminator also reshapes the reconstruction used for localization—an effect residual-only training lacks.
  • The recipe is modality-agnostic: it reports top results on MRI, ultrasound, and chest X-ray with the same two-stage training.
  • Cross-dataset evaluation on WMH after training on BraTS2021 suggests the learned shift signal is not limited to tumor-specific reconstruction.

Where Pith is reading between the lines

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

  • The image-level signal is only defined on the shift distribution induced by the fixed masking policy; a pathology that the diffusion model faithfully reconstructs would still be scored as normal. The authors flag exactly this transfer risk in their conclusion.
  • The same reconstruction-shift discrimination idea could be lifted out of diffusion: any controllable normal-image corruption, such as GAN inpainting or flow-based editing, could generate paired positives for the classifier.
  • A testable extension is to make the masking policy adaptive—selecting mask location and size per image—or to combine the classifier score with the residual map at inference, since the two signals fail in opposite directions.
  • Performance on the low-data BUSI setting (32 training images) suggests the classifier signal can be learned from very little normal data, but this is a stress test, not evidence of broad generalization.

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 / 6 minor

Summary. The paper proposes DMD, an unsupervised medical anomaly detection framework with two components: a latent diffusion restoration branch and a learned image-level discriminator. A VQ-VAE first learns a quantized latent space from normal images. In the second stage, random square masks corrupt selected latent positions, a latent diffusion denoiser reconstructs the masked latents, and the decoded reconstructions are paired with original normal images to train a binary classifier that distinguishes originals from reconstruction-shift counterparts. The diffusion and classification losses are optimized jointly, with classifier gradients flowing through the fixed decoder to the denoiser. At inference, the classifier is applied directly to the test image to produce an image-level anomaly score, while overlapping masked one-step restorations yield residual maps aggregated into a pixel-level anomaly map. Experiments on BraTS2021, ATLAS 2.0, WMH, BUSI, and VinDr-CXR compare DMD with classical and diffusion baselines, and the paper reports image-level and pixel-level metrics, ablations, and parameter sensitivity analyses.

Significance. If the performance holds up, DMD is a relevant contribution to unsupervised medical anomaly detection: it introduces a learned image-level signal derived entirely from normal training data, complements residual-based localization, and is evaluated across multiple modalities (brain MRI, ultrasound, chest X-ray) with normal-only threshold calibration. The ablation study isolates the contributions of the classifier, discriminative feedback, VQ pretraining, masking, and latent diffusion. However, the paper does not release code, provides no statistical significance tests, and leaves the central transfer assumption of the image-level score unquantified, so the strength of the empirical claims is currently limited.

major comments (3)
  1. [§4.2, Table 2, §5] The only cross-dataset/cross-pathology evaluation is WMH, but Table 2 reports only APpix and Dice for WMH; no image-level AUC/AP/F1 is given. The paragraph introducing WMH claims that DMD 'can generalize to pathology-level distribution shifts,' and the conclusion states that DMD is 'not limited to tumor-specific reconstruction.' Since the paper's central novelty is the learned image-level reconstruction-shift score, this claim is not supported for the image-level signal under domain shift. Please report image-level metrics on WMH (and ideally ATLAS 2.0) or restrict the generalization claim to pixel-level localization.
  2. [§3.6, Eq. (11)] The image-level score is produced by applying the classifier f_psi directly to a raw test image x*, even though f_psi was trained to separate original normal images from reconstructions produced by the masked diffusion process. The classifier therefore must transfer from synthetic reconstruction shifts (random masks applied to normal images) to arbitrary real pathologies, a distribution shift that the authors acknowledge in the conclusion but do not quantify. Without a score-distribution analysis (e.g., histograms or calibration of S_img for normal vs abnormal images on each dataset, and specifically on WMH), the mechanism by which the classifier separates real abnormalities is an unverified assumption. Please add such an analysis or temper the central claim.
  3. [§4.2, Tables 1-2] The claim that DMD 'achieves the best overall performance' is based on point estimates with mean±std over five runs, but no paired significance tests or confidence intervals are reported. Some comparisons are within one standard deviation, e.g., BUSI APpix: DMD 71.3±1.48 vs Dif-fuse 70.6±1.33, and BUSI Dice: DMD 72.4±1.30 vs Dif-fuse 71.8±1.36. To support the 'best overall' claim, please provide paired statistical tests (e.g., bootstrap confidence intervals or Wilcoxon signed-rank tests across runs) for the key comparisons, or report which differences are not statistically significant.
minor comments (6)
  1. [Table 4] The WMH row reads '1102D multi-slice FLAIR scans' in the text and '110' in the table; please clarify that this is 110 2D slices (or correct the value) to avoid ambiguity.
  2. [§4.2] Typo: 'DMD shows a advantage' should be 'DMD shows an advantage.'
  3. [Eq. (14)] The constant delta is introduced but its value is not given; please specify the numerical value used.
  4. [§4.1, Appendix C] The parameter-selection section states that thresholds and hyperparameters are selected on a validation split using only normal data, but the construction of this validation split is not described. This is especially relevant for BUSI, which uses only 32 normal training images; please clarify how the split is made and how stability is ensured.
  5. [Reproducibility] No code or trained models are released. Given the complexity of the two-stage training and the many hyperparameters, a public implementation would substantially improve reproducibility.
  6. [§4.2, Table 1] The text refers to MAD-AD as the strongest competitor, but on BraTS2021 the second-best AUC is THORSimplex (86.1 vs MAD-AD 85.7). Please state the strongest competitor per dataset or clarify that MAD-AD is strongest on average.

Circularity Check

0 steps flagged

No circularity: central claim is external benchmark performance; the learned image-level score is a self-supervised proxy, not a fitted input.

full rationale

The paper's central claim is empirical: DMD achieves the best overall performance on five external medical imaging benchmarks. The image-level anomaly score is produced by a classifier trained on a self-supervised task—distinguishing original normal images from their masked-diffusion reconstructions—with labels 0/1 generated entirely from normal training data. Equation (11) defines the test score as f_psi(x*), the classifier output; nothing in this definition reduces the target anomaly label to the training labels. The classifier is then evaluated on held-out normal and abnormal test images, and thresholds and hyperparameters are selected using only normal training/validation data, not test labels. The conclusion explicitly acknowledges that the reconstruction-shift proxy may not transfer across unseen pathologies and acquisition conditions; this is an external-validity caveat, not a circular derivation. The ablations compare variants against external metrics rather than against fitted values, and no fitted parameter is renamed as a prediction. Self-citations, if any, are to published benchmark protocols and normative metrics, which are independent external artifacts rather than load-bearing circular support. Therefore the derivation chain is self-contained and no circularity is present.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The ledger shows that DMD introduces no new physical or conceptual entities. Its claims rest on standard deep learning machinery plus a domain assumption about the transferability of reconstruction-shift features, which the authors partially acknowledge.

free parameters (4)
  • lambda (classification loss weight) = 0.1
    Tuned on validation split; sensitivity analysis (Fig. 7) shows best Dice at lambda=0.1.
  • mask area ratio range [rho_min, rho_max] = not stated; best around 900-1200 latent pixels
    Selected on validation; described as uniform range in Sec. 3.3 and analyzed in Fig. 6b.
  • inference restoration step t* = not stated; best around 500-600
    Fixed for inference, selected on validation; sensitivity analysis Fig. 6a.
  • inference mask side length l_M and stride s_M = not stated
    Selected on validation (Appendix C) and fixed for testing; no exact values given.
axioms (4)
  • domain assumption Reconstruction residuals are informative for anomaly localization
    The pixel-level anomaly map is derived directly from residuals between input and reconstruction (Eq. 13-14). If residuals are dominated by noise or normal variation, localization fails.
  • domain assumption The VQ-VAE latent space captures sufficient normal anatomy
    All masking and diffusion operate on the quantized latent grid; if the representation discards diagnostically relevant details, both classifier and residual maps suffer.
  • domain assumption Reconstruction-shift discrimination generalizes from synthetic masked shifts to real pathology
    The image-level score assumes that pathological images produce reconstruction shifts resembling those learned from normal-image reconstructions. The authors flag this as a limitation in the conclusion.
  • standard math Standard diffusion and VQ-VAE training objectives
    The method relies on the standard DDPM noise schedule and VQ loss as given in Eqs. 4-6.

pith-pipeline@v1.3.0-alltime-deepseek · 17282 in / 12020 out tokens · 115660 ms · 2026-08-05T00:57:15.011822+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Reconstruction-Shift Discrimination via Mask-Guided Latent Diffusion for Medical Anomaly Detection." pith.science (2026). https://pith.science/paper/74SX6J2R

@misc{pith2026260800444,
  author       = {Pith},
  title        = {Pith review of: Reconstruction-Shift Discrimination via Mask-Guided Latent Diffusion for Medical Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/74SX6J2R}},
  note         = {Machine review of arXiv:2608.00444}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Unsupervised medical anomaly detection learns normal anatomical patterns from healthy training images and identifies deviations at test time. Reconstruction-based and diffusion-based methods commonly use the difference between an input image and its reconstruction as anomaly evidence. However, this residual can be ambiguous. Expressive models may preserve pathological structures, while benign anatomical variation, imaging noise, and acquisition differences may also produce large reconstruction errors. We propose discriminative mask-guided diffusion (DMD), a medical anomaly detection framework that complements residual-based localization with reconstruction-shift discrimination. DMD first learns a compact quantized latent representation of normal images. Localized masks then perturb selected latent regions, and a latent diffusion model reconstructs the perturbed representations. The resulting reconstructions are paired with their original normal images to define a self-supervised classification task. At inference, the classifier provides a learned image-level anomaly score, while the residual between the input and its diffusion-based reconstruction yields a pixel-level anomaly map. Experiments on five datasets spanning brain MRI, breast ultrasound, and chest radiography show that DMD achieves the best overall performance among the state-of-the-art baseline methods.

Figures

Figures reproduced from arXiv: 2608.00444 by Jinyu Cai, See-kiong Ng, Yi Bin, Yibo Wan, Yunhe Zhang.

Figure 1
Figure 1. Figure 1: Overview of DMD. In Stage I, a vector-quantized encoder–decoder learns a compact normal latent representation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of anomaly localization [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: T-SNE visualization of feature embeddings ex [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: More visual comparison of anomaly detection re [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Parameter sensitivity analysis of DMD. (a) Dice [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Trade-off λ Parameter sensitivity analysis on BraTS2021 Across all timestep settings, DMD consistently outperforms cDDPM and MAD-AD, with the best performance observed around t = 500–600. Figure 6b evaluates the effect of mask region size on BraTS2021, ATLAS 2.0, and BUSI. Across datasets, performance generally improves when the mask size increases from small regions to a moderate range, indicating that lo… view at source ↗

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

39 extracted references · 34 canonical work pages · 1 internal anchor

  1. [1]

    Multimedia Tools and Applications , volume =

    Abhisheka, Barsha and Biswas, Saroj Kumar and Purkayastha, Biswajit and Das, Dolly and Escargueil, Alexandre , title =. Multimedia Tools and Applications , volume =. 2024 , type =

    work page 2024

  2. [2]

    Procedia Computer Science , volume=

    Masked autoencoders for unsupervised anomaly detection in medical images , author=. Procedia Computer Science , volume=. 2023 , publisher=

    work page 2023

  3. [3]

    European Conference on Computer Vision , pages=

    Natural synthetic anomalies for self-supervised anomaly detection and localization , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    work page 2022

  4. [4]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Draem-a discriminatively trained reconstruction embedding for surface anomaly detection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  5. [5]

    Medical Image Analysis , volume=

    MedIAnomaly: A comparative study of anomaly detection in medical images , author=. Medical Image Analysis , volume=. 2025 , publisher=

    work page 2025

  6. [6]

    Pattern Recognition , pages=

    Multi-model co-training for medical image segmentation with limited annotation , author=. Pattern Recognition , pages=. 2025 , publisher=

    work page 2025

  7. [7]

    ArXiv:2107.02314 , year =

    Baid, Ujjwal and Ghodasara, Satyam and Mohan, Suyash and Bilello, Michel and Calabrese, Evan and Colak, Errol and Farahani, Keyvan and Kalpathy-Cramer, Jayashree and Kitamura, Felipe C and Pati, Sarthak , title =. ArXiv:2107.02314 , year =

    Pith/arXiv arXiv

  8. [8]

    Medical Image Analysis , volume =

    Baur, Christoph and Denner, Stefan and Wiestler, Benedikt and Navab, Nassir and Albarqouni, Shadi , title =. Medical Image Analysis , volume =. 2021 , type =

    work page 2021

  9. [9]

    Deep autoencoding models for unsupervised anomaly segmentation in brain MR images , author=. Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 4th International Workshop, BrainLes 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Revised Selected Papers, Part I 4 , pages=. 2019 , organization=

    work page 2018

  10. [10]

    Medical Imaging with Deep Learning , pages=

    Patched diffusion models for unsupervised anomaly detection in brain mri , author=. Medical Imaging with Deep Learning , pages=. 2024 , organization=

    work page 2024

  11. [11]

    International Workshop on Machine Learning in Medical Imaging , pages=

    Unsupervised anomaly detection in medical images using masked diffusion model , author=. International Workshop on Machine Learning in Medical Imaging , pages=. 2023 , organization=

    work page 2023

  12. [12]

    Medical Image Analysis , volume =

    Schlegl, Thomas and Seeböck, Philipp and Waldstein, Sebastian M and Langs, Georg and Schmidt-Erfurth, Ursula , title =. Medical Image Analysis , volume =. 2019 , type =

    work page 2019

  13. [13]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , publisher =

    Wolleb, Julia and Bieder, Florentin and Sandkühler, Robin and Cattin, Philippe C , title =. International Conference on Medical Image Computing and Computer-Assisted Intervention , publisher =. 2022 , type =

    work page 2022

  14. [14]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Zhang, Xinyi and Li, Naiqi and Li, Jiawei and Dai, Tao and Jiang, Yong and Xia, Shu-Tao , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2023 , type =

    work page 2023

  15. [15]

    Medical Imaging with Deep Learning , year=

    Generalizing unsupervised anomaly detection: towards unbiased pathology screening , author=. Medical Imaging with Deep Learning , year=

    work page

  16. [16]

    Medical Imaging 2019: Image Processing , volume=

    Unsupervised brain lesion segmentation from MRI using a convolutional autoencoder , author=. Medical Imaging 2019: Image Processing , volume=. 2019 , organization=

    work page 2019

  17. [17]

    IXI Dataset - Brain Development , year =

    work page

  18. [18]

    Scientific data , volume=

    A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms , author=. Scientific data , volume=. 2022 , publisher=

    work page 2022

  19. [19]

    2022 , publisher=

    Data of the White Matter Hyperintensity (WMH) Segmentation Challenge , author=. 2022 , publisher=

    work page 2022

  20. [20]

    Zeitschrift Fuer Medizinische Physik , volume=

    An overview of deep learning in medical imaging focusing on MRI , author=. Zeitschrift Fuer Medizinische Physik , volume=. 2019 , publisher=

    work page 2019

  21. [21]

    Scientific Reports , volume=

    A hybrid explainable model based on advanced machine learning and deep learning models for classifying brain tumors using MRI images , author=. Scientific Reports , volume=. 2025 , publisher=

    work page 2025

  22. [22]

    MAD-AD: Masked Diffusion for Unsupervised Brain Anomaly Detection

    MAD-AD: Masked Diffusion for Unsupervised Brain Anomaly Detection , author=. ArXiv:2502.16943 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Diffusion models with implicit guidance for medical anomaly detection , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2024 , organization=

    work page 2024

  24. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  25. [25]

    Advances in Neural Information Processing Systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in Neural Information Processing Systems , volume=

  26. [26]

    Computers in Biology and Medicine , volume=

    Guided reconstruction with conditioned diffusion models for unsupervised anomaly detection in brain mris , author=. Computers in Biology and Medicine , volume=. 2025 , publisher=

    work page 2025

  27. [27]

    IEEE Transactions on Medical Imaging , year=

    Diffusion models for counterfactual generation and anomaly detection in brain images , author=. IEEE Transactions on Medical Imaging , year=

    work page

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [29]

    Data in brief , volume=

    Dataset of breast ultrasound images , author=. Data in brief , volume=. 2020 , publisher=

    2020

  30. [30]

    Frontiers in Oncology , volume=

    The global burden of breast cancer in women from 1990 to 2030: assessment and projection based on the global burden of disease study 2019 , author=. Frontiers in Oncology , volume=

    work page 1990

  31. [31]

    Scientific Data , volume=

    VinDr-CXR: An open dataset of chest X-rays with radiologist’s annotations , author=. Scientific Data , volume=. 2022 , publisher=

    work page 2022

  32. [32]

    ArXiv:1511.05644 , year=

    Adversarial autoencoders , author=. ArXiv:1511.05644 , year=

    Pith/arXiv arXiv

  33. [33]

    Nature Communications , volume=

    Evaluating normative representation learning in generative AI for robust anomaly detection in brain imaging , author=. Nature Communications , volume=. 2025 , publisher=

    work page 2025

  34. [34]

    Medical Image Analysis , volume=

    Unsupervised brain imaging 3D anomaly detection and segmentation with transformers , author=. Medical Image Analysis , volume=. 2022 , publisher=

    work page 2022

  35. [35]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Fast unsupervised brain anomaly detection and segmentation with diffusion models , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2022 , organization=

    work page 2022

  36. [36]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Itermask 2: Iterative unsupervised anomaly segmentation via spatial and frequency masking for brain lesions in mri , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2024 , organization=

    work page 2024

  37. [37]

    Medical Image Analysis , pages=

    IterMask3D: Unsupervised anomaly detection and segmentation with test-time iterative mask refinement in 3D brain MRI , author=. Medical Image Analysis , pages=. 2025 , publisher=

    work page 2025

  38. [38]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Cutpaste: Self-supervised learning for anomaly detection and localization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  39. [39]

    Medical image analysis , volume=

    The role of noise in denoising models for anomaly detection in medical images , author=. Medical image analysis , volume=. 2023 , publisher=

    work page 2023