Pith. sign in

REVIEW 4 major objections 4 minor 35 references

This paper claims that enforcing D4 rotation/reflection symmetry in diffusion models stabilizes anomaly ranking in computational cytology and improves top-K abnormal-cell retrieval.

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

2026-08-01 02:13 UTC pith:UACOYOS6

load-bearing objection A well-motivated application of equivariant diffusion to cytology, but the main empirical claim is confounded by a noise-type change and needs a proper ablation. the 4 major comments →

arxiv 2607.25503 v1 pith:UACOYOS6 submitted 2026-07-28 cs.CV

Group Equivariant Diffusion for Anomaly Detection in Computational Cytology

Swarnadip Chatterjee , Ssharvien Kumar Sivakumar , Anirban Mukhopadhyay This is my paper
classification cs.CV
keywords computational cytologyanomaly detectiondiffusion modelsD4-equivariancedihedral groupwhole-slide imagingtop-K retrievalsingle-cell morphology
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 argues that in cell-centric cytology, a single-cell patch's diagnostic class is unchanged by quarter-turn rotations and mirror flips, yet ordinary diffusion anomaly detectors treat each transformed view as a separate input, so anomaly scores wobble with orientation. It proposes building that dihedral (D4) symmetry into a diffusion model both in the denoiser architecture and in the sampling process, via equivariant noise coupling and frame averaging. On two public cytology datasets, the resulting D4-equivariant models beat non-equivariant generative baselines, a one-class method, and a multiple-instance-learning scorer, and they substantially lower score variance across orientations. If correct, the central insight is that encoding a known biological invariance removes nuisance variability and makes the top of the ranked list more reliable under extreme class imbalance.

Core claim

The central claim is that enforcing D4 equivariance — in-plane rotations by multiples of 90 degrees plus reflections — throughout partial-diffusion anomaly detection yields transformation-consistent pseudo-healthy reconstructions and more stable anomaly scores, and that this improves both global separation (AUC) and top-K retrieval of abnormal cells on bone marrow and peripheral blood smears. The paper reports that on the MLL dataset, the D4 frame-averaged plus equivariant-noise variant raises TP400 from 57 (AnoDDPM) to 78 with paired-bootstrap p=0.006, while on AML LMU the D4-equivariant attention U-Net reaches the best AUC (0.720), though the gain over BerDiff is not statistically conclusi

What carries the argument

D4-equivariant diffusion combines three mechanisms: a D4-equivariant U-Net denoiser (group-tied convolution kernels, optionally a D4-equivariant attention block with shared group-transformed relative positional biases); equivariant noise coupling, which applies the same base noise sample to every rotated/flipped view via the group action; and frame averaging, which averages residuals and scores over the D4 orbit. The D4 group is the symmetry group of the square — rotations by 0, 90, 180, 270 degrees and four reflections. The machinery works by making the stochastic reconstruction trajectory pathwise equivariant, so patches that differ only by a D4 transform receive nearly identical anomaly s

Load-bearing premise

The load-bearing premise is that the diagnostic class of a centered single-cell patch is unchanged by 90-degree rotations and mirror flips, so enforcing this symmetry cannot remove discriminative morphology; the paper states this as class invariance but never verifies orientation invariance of abnormal cells empirically.

What would settle it

Train an equivariant and a non-equivariant anomaly detector on normal patches, then evaluate on a held-out set of abnormal cells known or labeled to have orientation-dependent morphology, such as polarized granules or asymmetric nuclear blebs. If the D4-equivariant model's AUC or top-K retrieval on that subset is materially worse than the non-equivariant baseline's, the central claim that enforcing D4 cannot lose signal is refuted. A synthetic experiment that injects oriented abnormal textures into normal images would settle it directly.

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

If this is right

  • On MLL, D4 FA+EN raises true positives among the top 400 ranked patches from 57 to 78 with p=0.006, meaning a budgeted reviewer sees more actual abnormal cells earlier.
  • Score variance across rotations and flips roughly halves: mean SD drops from 0.00375 to 0.00206, so the ranking order is stable under viewing transforms.
  • The D4-equivariant attention U-Net achieves the best AUC (0.720) and TP400 (325) on the AML LMU blood-smear dataset, showing the approach transfers across bone marrow and peripheral blood.
  • The proof that a D4-averaged denoiser has no higher DDPM risk than a plain denoiser justifies symmetry-respecting averaging whenever the normal distribution is D4-invariant.

Where Pith is reading between the lines

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

  • An extension the paper does not pursue: the same argument applies to any cell-centric assay with planar orientation invariance, such as fluorescence microscopy or Pap-smear patches, where D4-equivariant diffusion could be dropped in without retraining the inductive bias.
  • The normal class is modeled as a single homogeneous distribution, but cytology normals are often mixtures of subtypes; if the normal mixture is multimodal, a D4-equivariant model may over-smooth, and per-subtype reconstruction error would clarify where the assumption bites.
  • A stress test the paper leaves implicit: measure per-orientation AUC on abnormal cells with oriented morphology; if any abnormal subtype is orientation-dependent, equivariance could erase the signal, and the paper's own score-variance reduction would hide that loss.

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

4 major / 4 minor

Summary. The paper proposes a D4-equivariant diffusion framework for unsupervised anomaly detection in computational cytology. It enforces rotation/reflection symmetry both architecturally (D4-equivariant U-Net, optionally with attention) and at inference (equivariant noise and frame averaging), and evaluates patch-level anomaly ranking on two public cytology datasets, MLL bone marrow and AML LMU. The main claims are that the D4 variants improve AUC and top-K true-positive retrieval over non-equivariant generative baselines, deep one-class, and MIL methods, while substantially reducing anomaly-score variance under D4 transforms. Statistical significance is reported via paired bootstrap tests, with strong gains on MLL and numerically consistent but statistically inconclusive gains on AML LMU.

Significance. If the central claim were fully established, the work would be a useful contribution: it addresses a domain-specific symmetry that is plausible for centered single-cell patches, provides a concrete instantiation of equivariant diffusion for cytology, and evaluates on publicly available datasets with code. The manuscript also includes paired bootstrap uncertainty quantification and an explicit discussion of the statistically inconclusive AML result, which is commendable. However, the current evidence for the central claim is weakened by a confounded baseline comparison, an internal inconsistency between the algorithm description and the reported variants, and a headline invariance result that is partly forced by construction. These issues are addressable, but they are load-bearing for the paper's main message.

major comments (4)
  1. [§3.2, Algorithm 1, Table 1] The primary positive result on MLL (AnoDDPM AUC 0.632/TP400 57 vs. FA+EN AUC 0.684/TP400 78, p<1e-3 / p=0.006) changes two factors at once. The AnoDDPM baseline is defined in §3.2 as a 'non-equivariant U-Net, simplex noise, partial diffusion', whereas Algorithm 1 (lines 4, 11, 17) specifies Gaussian noise for all proposed variants. There is no non-equivariant Gaussian partial-diffusion baseline. The FA-only result (MLL AUC 0.612, TP400 46) shows that Gaussian noise plus frame averaging is not uniformly better, but it does not rule out an interaction between noise type and the FA+EN or architectural equivariance coupling. Please add an ablation: a non-equivariant U-Net trained with Gaussian noise and identical partial-diffusion settings (same t*, T, schedule, hyperparameters), and include the same paired bootstrap comparisons against AnoDDPM. Without this control, the observed gains canno
  2. [§2.2 vs. Algorithm 1] Algorithm 1 has no EN-only branch. When FA is false (lines 16–18), the EN flag is ignored; when FA is true, EN only modifies the noise inside the orbit loop. Yet §2.2 states that the D4-equivariant U-Net variants 'typically use EN-only'. Consequently, the Table 1 rows 'D4 conv.' and 'D4 conv.+attn.' and the §3.4 invariance numbers for the 'D4 U-Net without attention' are not reproducible from the pseudocode. Please specify the exact inference mode used for each reported row, and either implement EN-only in Algorithm 1 or clarify that the architectural variants simply run line 17 with a single equivariant sample, which is not 'equivariant noise coupling' in the sense of the orbit loop.
  3. [§3.4, 'Anomaly-score invariance under D4 transforms'] For the FA+EN variant, the reported reduction in score variance is a construction artifact. Algorithm 1 lines 8–15 define the final score as the average of S_g over the D4 orbit, so S_FA(g·x0)=S_FA(x0) for every g by definition, up to numerical precision. Reporting a per-image SD for FA+EN as an empirical 'reduction' overstates the finding. The meaningful invariance measurements are for the vanilla baseline and for any architectural variant that does not average over the orbit. Please report those separately, and state explicitly that the FA+EN result is enforced by construction.
  4. [§2.3] The theoretical guarantees assume D4-invariant class-conditionals, p(x|y)=p(g·x|y) for y∈{0,1}, but the paper does not empirically verify this assumption for the abnormal classes. If any abnormal morphology is orientation-dependent (e.g., polarized blasts or Auer-rod-like structures), enforcing D4 equivariance could remove discriminative signal. This is not a circularity objection but a correctness-risk concern. Please provide a direct test, e.g., compare a non-equivariant patch classifier's accuracy on original versus rotated/flipped abnormal patches, or measure how well the abnormal class-conditional distribution is preserved under D4 transforms. This would support the central motivation that D4 invariance is biologically appropriate for the specific cell types studied.
minor comments (4)
  1. [Table 1] The AML LMU AnoDDPM row reads '31116283'; this likely should be '311, 162, 83' for TP400, TP200, TP100. Please fix the formatting.
  2. [§3.1] Minor naming/formatting: 'AMLLMUCytomorphology' should be 'AML LMU Cytomorphology' or similar.
  3. [§3.4] The phrase 'D4-equivariant AnoDDPM variants' may confuse the reader, since the baselines are defined with simplex noise while the proposed models use Gaussian noise. Please clarify the terminology.
  4. [§3.4] Please report the number of bootstrap resamples and the random seeds used for the paired tests, so the p-values are reproducible.

Circularity Check

1 steps flagged

The headline stability result is definitional: frame averaging over the D4 orbit makes the FA+EN score group-invariant by construction, so the reported reduction in score variance is a restatement of the estimator, not an empirical discovery. The AUC/TP improvements remain empirical but are partly confounded with a noise-distribution change.

specific steps
  1. self definitional [Algorithm 1 (lines 8–15) + Section 2.2 + Section 3.4 'Anomaly-score invariance under D4 transforms']
    "When desired, we further apply frame averaging (FA), averaging residuals and scores over the D4 orbit to stabilize A(x0) and S(x0). ... S F A(x0) = 1 |G| P g∈G Sg;returnS F A(x0) ... Vanilla AnoDDPM yields (0.00375,0.00343,0.01036), whereas the D4 FA+EN variant reduces this to (0.00206,0.00161,0.00515)."

    For the FA+EN variant, the score of a patch is defined as the mean of the per-group scores over its full D4 orbit. With the equivariant-noise coupling in Algorithm 1, evaluating a rotated/flipped version with the same base noise permutes the terms of the same orbit sum, so S_FA(h·x0)=S_FA(x0) identically (in expectation/exactly with shared randomness). The measured 'substantially reduced' score variance is therefore a consequence of the averaging definition, not an empirical property learned from data. Reporting it as a benefit of D4-equivariant diffusion inserts the invariance by construction and then presents it as a discovered outcome.

full rationale

The theoretical guarantees in Section 2.3 are self-contained, parameter-free inequalities under stated D4-invariance assumptions, so they are not circular. The AUC/TP400 comparisons on MLL and AML LMU are empirical evaluations against external baselines and are not derived from the method's definition. However, the paper's headline 'substantially reducing score variance across rotations and flips' reduces directly to the frame-averaging operation in Algorithm 1, which is self-definitional: averaging over the D4 orbit removes orientation-dependent score variation by construction. In addition, the main statistically significant MLL comparison (FA+EN vs AnoDDPM) simultaneously changes equivariance and noise distribution (Gaussian vs simplex), so the equivariance contribution is not isolated; this is an experimental confound rather than a formal circularity. No load-bearing self-citation chain is present. Overall, the central empirical ranking claim retains independent content, but one of the paper's two advertised advantages is true by construction, yielding partial circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on the D4 class-invariance domain assumption, several undisclosed hyperparameters (t*, T, architecture, training), and a potential noise-type confound. No genuinely new entities are introduced; the 'invented' components (equivariant noise, frame averaging) are re-appropriations of refs [22,35].

free parameters (5)
  • partial diffusion step t* = not reported
    The number of forward/reverse steps used at inference, a key hyperparameter for reconstruction-based anomaly scores; not disclosed in the paper, likely tuned per dataset.
  • total diffusion steps T and noise schedule = not reported
    Standard DDPM schedule with T steps; T not specified, so the exact process is unreproducible from the text.
  • model architecture hyperparameters (channels, depth, attention, group order) = not reported
    D4-equivariant U-Net details are not specified; the capacity of the equivariant models is not controlled against baselines.
  • noise type (Gaussian vs simplex) = Gaussian (method), simplex (AnoDDPM baseline)
    The main baseline uses simplex noise while the proposed variants are described with Gaussian noise, so the equivariance comparison may be confounded by noise type.
  • training hyperparameters (batch size, learning rate, epochs, seeds) = not reported
    No training configuration is given; stochastic results are reported without across-seed variance.
axioms (4)
  • domain assumption Class invariance: p(x|y) = p(g·x|y) for all g in D4 and y in {0,1}, i.e., the diagnostic class of a centered single-cell patch is unchanged by 90-degree rotations and reflections.
    Introduced in Section 1 and formalized in Section 2.3; never empirically tested on these datasets. If false (oriented morphology), equivariance may erase diagnostic signal.
  • domain assumption The normal training patches (LYT for MLL; mature leukocytes for AML) are drawn i.i.d. from a single normal distribution p_n.
    Training and test split of the normal class assumes homogeneity; atypical benign/reactive cells within the 'normal' labels could act as hidden anomalies.
  • standard math DDPM training with mean-squared noise prediction (L(θ) = E||ϵ - f_θ(x_t, t)||^2) trains a good denoiser for normal patches.
    Standard DDPM formulation from Ho et al., used in Section 2.1; accepted background.
  • standard math Jensen's inequality yields the D4-averaged denoiser cannot increase DDPM risk.
    Section 2.3; a standard convexity argument that holds for the squared-error loss.

pith-pipeline@v1.3.0-alltime-deepseek · 8356 in / 16269 out tokens · 153022 ms · 2026-08-01T02:13:48.700564+00:00 · methodology

0 comments
read the original abstract

Computational cytology on whole-slide images is challenging because malignant cells are rare, heterogeneous, and annotated slides are scarce. Anomaly detection frameworks can be trained on normal slide-negative patches and then applied at test time to flag abnormal patches in held-out slides. Most unsupervised anomaly detection approaches including generative ones (GAN-based and diffusion-based), are tuned to organ-level imaging and require large curated datasets. In cytology the signal is cell-centric: rotating or flipping a single-cell patch does not change its diagnostic class, yet standard diffusion models treat transformed views as distinct inputs, leading to transformation-dependent reconstructions and unstable anomaly scores. We propose a D4-equivariant diffusion framework that enforces rotation and reflection symmetry both architecturally, via a D4-equivariant U-Net, and at inference, via equivariant noise coupling and (optionally) frame averaging. This alignment with biological invariance yields transformation-consistent pseudo-healthy reconstructions and more stable anomaly ranking under symmetry. On two publicly available cytology datasets of bone marrow and peripheral blood smears, our D4-equivariant diffusion models achieve higher AUC and retrieve more abnormal cells in the top K predictions than non-equivariant generative baselines, a deep one-class, and a multiple instance learning based method, while substantially reducing score variance across rotations and flips. Code is available at https://swchmida.github.io/D4diffCyto/.

Figures

Figures reproduced from arXiv: 2607.25503 by Anirban Mukhopadhyay, Ssharvien Kumar Sivakumar, Swarnadip Chatterjee.

Figure 1
Figure 1. Figure 1: Overview of the proposed D4-equivariant diffusion framework. Left: Cayley diagram of the dihedral group D4 acting on a single-cell patch. Top-right: training with a D4-equivariant U-Net with equivariant noise (EN) coupling. Bottom-right: Inference using partial diffusion reconstruction at test time and an example D4-tied convolution kernel. and weakly supervised multiple instance learning (MIL) based slide… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of AnoDDPM vs. D4-equivariant AnoDDPM on MLL bone marrow (left panel) and AML LMU (right panel). For each method we show in￾puts (row 1 ), reconstructions (rows 2 and 4 ), and qualitative reconstruction-deviation overlays with patch-level anomaly-score MSE values (rows 3 and 5 ). 3.4 Ablation studies Inference-induced vs. architectural D4 variants. We ablate four D4 vari￾ants ( [PIT… 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

35 extracted references · 1 linked inside Pith

  1. [1]

    Avicenna journal of medicine1(01), 18–28 (2011)

    Al-Abbadi, M.A.: Basics of cytology. Avicenna journal of medicine1(01), 18–28 (2011)

    2011

  2. [2]

    In: IPMI

    Beizaee, F., Lodygensky, G., Desrosiers, C., Dolz, J.: Mad-ad: Masked diffusion for unsupervised brain anomaly detection. In: IPMI. pp. 139–153. Springer (2025)

    2025

  3. [3]

    Computational and mathematical methods in medicine2014(1), 842037 (2014)

    Bengtsson, E., Malm, P.: Screening for cervical cancer using automated analysis of pap-smears. Computational and mathematical methods in medicine2014(1), 842037 (2014)

    2014

  4. [4]

    Machine Learning for Biomedical Imaging (2025)

    Bercea,C.I.,Cattin,P.C.,Schnabel,J.A.,Wolleb,J.:Denoisingdiffusionmodelsfor anomaly localization in medical images. Machine Learning for Biomedical Imaging (2025)

    2025

  5. [5]

    arXiv preprint arXiv:2305.19643 (2023)

    Bercea, C.I., Neumayr, M., Rueckert, D., Schnabel, J.A.: Mask, stitch, and re- sample: Enhancing robustness and generalizability in anomaly detection through automatic diffusion models. arXiv preprint arXiv:2305.19643 (2023)

    Pith/arXiv arXiv 2023

  6. [6]

    In: MICCAI

    Bercea, C.I., Wiestler, B., Rueckert, D., Schnabel, J.A.: Diffusion models with im- plicit guidance for medical anomaly detection. In: MICCAI. pp. 211–220. Springer (2024)

    2024

  7. [7]

    In: ICPR

    Chatterjee, S., Göksel, O., Sladoje, N., Lindblad, J.: Detection of extremely sparse key instances in whole slide cytology images via self-supervised one-class represen- tation learning. In: ICPR. pp. 408–421. Springer (2024)

    2024

  8. [8]

    Nature communications12(1), 5639 (2021)

    Cheng, S., Liu, S., Yu, J., Rao, G., Xiao, Y., Han, W., Zhu, W., Lv, X., Li, N., Cai, J., et al.: Robust whole slide image analysis for cervical cancer screening using deep learning. Nature communications12(1), 5639 (2021)

    2021

  9. [9]

    In: IEEE CVPR Workshops (2019)

    Chidester, B., Ton, T.V., Tran, M.T., Ma, J., Do, M.N.: Enhanced rotation- equivariant u-net for nuclear segmentation. In: IEEE CVPR Workshops (2019)

    2019

  10. [10]

    Blood, The Journal of the American Society of Hematology117(23), 6267–6276 (2011)

    Coustan-Smith, E., Song, G., Clark, C., Key, L., Liu, P., Mehrpooya, M., Stow, P., Su, X., Shurtleff, S., Pui, C.H., et al.: New markers for minimal residual disease detection in acute lymphoblastic leukemia. Blood, The Journal of the American Society of Hematology117(23), 6267–6276 (2011)

    2011

  11. [11]

    ACM Computing Surveys (CSUR)54(7), 1–37 (2021)

    Fernando, T., Gammulle, H., Denman, S., Sridharan, S., Fookes, C.: Deep learning for medical anomaly detection–a survey. ACM Computing Surveys (CSUR)54(7), 1–37 (2021)

    2021

  12. [12]

    In: Medical Imaging with Deep Learning

    Fuchs, M., Sivakumar, S.K.R., Schöber, M., Woltering, N., Eich, M.L., Schweizer, L., Mukhopadhyay, A.: Harp: Unsupervised histopathology artifact restoration. In: Medical Imaging with Deep Learning. pp. 465–479. PMLR (2024)

    2024

  13. [13]

    In: ICML

    Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: ICML. pp. 2127–2136. PMLR (2018)

    2018

  14. [14]

    Medical Image Analysis84, 102691 (2023) 10 S

    Jiang, H., Zhou, Y., Lin, Y., Chan, R.C., Liu, J., Chen, H.: Deep learning for computational cytology: A survey. Medical Image Analysis84, 102691 (2023) 10 S. Chatterjee et al

    2023

  15. [15]

    International journal of molecular sciences21(3), 1054 (2020)

    Kruse, A., Abdel-Azim, N., Kim, H.N., Ruan, Y., Phan, V., Ogana, H., Wang, W., Lee, R., Gang, E.J., Khazal, S., et al.: Minimal residual disease detection in acute lymphoblastic leukemia. International journal of molecular sciences21(3), 1054 (2020)

    2020

  16. [16]

    Medical Image Analysis68, 101849 (2021)

    Lafarge, M.W., Bekkers, E.J., Pluim, J.P., Duits, R., Veta, M.: Roto-translation equivariant convolutional networks: Application to histopathology image analysis. Medical Image Analysis68, 101849 (2021)

    2021

  17. [17]

    Journal of the American Society of Cytopathology8(4), 230–241 (2019)

    Landau, M.S., Pantanowitz, L.: Artificial intelligence in cytopathology: a review of the literature and overview of commercial landscape. Journal of the American Society of Cytopathology8(4), 230–241 (2019)

    2019

  18. [18]

    In: IEEE ICCV

    Li, Y., Liu, Y., Xu, Y., Zhang, L., Xing, L., Huang, J.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: IEEE ICCV. pp. 14318–14327 (2021)

    2021

  19. [19]

    Medical Image Analysis93, 103088 (2024)

    Linmans, J., Raya, G., van der Laak, J., Litjens, G.: Diffusion models for out- of-distribution detection in digital pathology. Medical Image Analysis93, 103088 (2024)

    2024

  20. [20]

    In: MICCAI

    Liu, J., Mao, A., Niu, Y., Zhang, X., Gong, T., Li, C., Gao, Z.: Pamil: Prototype attention-based multiple instance learning for whole slide image classification. In: MICCAI. pp. 362–372. Springer (2024)

    2024

  21. [21]

    In: IEEE CVPR

    Liu, K., Zhu, W., Shen, Y., Liu, S., Razavian, N., Geras, K.J., Fernandez-Granda, C.: Multiple instance learning via iterative self-paced supervised contrastive learn- ing. In: IEEE CVPR. pp. 3355–3365 (2023)

    2023

  22. [22]

    In: AISTATS (2025)

    Lu, H., Szabados, S., Yu, Y.: Diffusion models under group transformations. In: AISTATS (2025)

    2025

  23. [23]

    Nature biomedical engineering5(6), 555–570 (2021)

    Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering5(6), 555–570 (2021)

    2021

  24. [24]

    The Can- cer Imaging Archive (TCIA) [Data set] (2021)

    Matek, C., Krappe, S., Münzenmayer, C., Haferlach, T., Marr, C.: An expert- annotated dataset of bone marrow cytology in hematologic malignancies. The Can- cer Imaging Archive (TCIA) [Data set] (2021)

    2021

  25. [25]

    The Cancer Imaging Archive (TCIA) [Data set] (2019)

    Matek, C., Schwarz, S., Marr, C., Spiekermann, K.: A single-cell morphological dataset of leukocytes from aml patients and non-malignant controls. The Cancer Imaging Archive (TCIA) [Data set] (2019)

    2019

  26. [26]

    In: MICCAI

    Naval Marimont, S., Siomos, V., Baugh, M., Tzelepis, C., Kainz, B., Tarroni, G.: Ensembled cold-diffusion restorations for unsupervised anomaly detection. In: MICCAI. pp. 243–253. Springer (2024)

    2024

  27. [27]

    Proceedings of the IEEE109(5), 756–795 (2021)

    Ruff, L., Kauffmann, J.R., Vandermeulen, R.A., Montavon, G., Samek, W., Kloft, M., Dietterich, T.G., Müller, K.R.: A unifying review of deep and shallow anomaly detection. Proceedings of the IEEE109(5), 756–795 (2021)

    2021

  28. [28]

    In: ICML

    Ruff, L., Vandermeulen, R.A., Görnitz, N., Deecke, L., Siddiqui, S.A., Binder, A., Müller, E., Kloft, M.: Deep one-class classification. In: ICML. pp. 4393–4402. PMLR (2018)

    2018

  29. [29]

    Medical image analysis79, 102464 (2022)

    Schirris, Y., Gavves, E., Nederlof, I., Horlings, H.M., Teuwen, J.: Deepsmile: Con- trastiveself-supervisedpre-trainingbenefitsmsiandhrdclassificationdirectlyfrom h&e whole-slide images in colorectal and breast cancer. Medical image analysis79, 102464 (2022)

    2022

  30. [30]

    Schlegl,T.,Seeböck,P.,Waldstein,S.M.,Langs,G.,Schmidt-Erfurth,U.:f-anogan: Fastunsupervisedanomalydetectionwithgenerativeadversarialnetworks.Medical image analysis54, 30–44 (2019)

    2019

  31. [31]

    Pattern Recognition Letters186, 64–70 (2024) Equivariant Diffusion for Anomaly Detection in Cytology 11

    Torpey, D., Klein, R.: Deepset simclr: Self-supervised deep sets for improved pathology representation learning. Pattern Recognition Letters186, 64–70 (2024) Equivariant Diffusion for Anomaly Detection in Cytology 11

    2024

  32. [32]

    In: MICCAI

    Wolleb, J., Bieder, F., Friedrich, P., Zhang, P., Durrer, A., Cattin, P.C.: Binary noise for binary tasks: Masked bernoulli diffusion for unsupervised anomaly detec- tion. In: MICCAI. pp. 135–145. Springer (2024)

    2024

  33. [33]

    In: MICCAI

    Wolleb, J., Bieder, F., Sandkühler, R., Cattin, P.C.: Diffusion models for medical anomaly detection. In: MICCAI. pp. 35–45. Springer (2022)

    2022

  34. [34]

    In: IEEE CVPR

    Wyatt, J., Leach, A., Schmon, S.M., Willcocks, C.G.: Anoddpm: Anomaly detec- tion with denoising diffusion probabilistic models using simplex noise. In: IEEE CVPR. pp. 650–656 (2022)

    2022

  35. [35]

    In: ICLR (2024),https://openreview.net/forum?id=FKksTayvGo

    Zhou, L., Lou, A., Khanna, S., Ermon, S.: Denoising diffusion bridge models. In: ICLR (2024),https://openreview.net/forum?id=FKksTayvGo

    2024