List of Papers Browse by Subject Areas Author List
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/.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2803_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/SwchMida/D4diffCyto
Link to the Dataset(s)
MLL dataset: https://doi.org/10.7937/6PPE-4020
AML LMU dataset: https://doi.org/10.7937/tcia.2019.36f5o9ld
BibTex
@InProceedings{ChaSwa_Group_MICCAI2026,
author = { Chatterjee, Swarnadip AND Sivakumar, Ssharvien Kumar AND Mukhopadhyay, Anirban},
title = { { Group Equivariant Diffusion for Anomaly Detection in Computational Cytology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a diffusion-based anomaly detection framework for computational cytology that explicitly enforces equivariance for cell images. In particular, the method incorporates a D4-equivariant U-Net architecture, equivariant noise coupling and frame averaging. The author further provides both theoretical arguments and empirical evidence demonstrating the effectiveness of the proposed approach.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- Strong logical grounding: The author provides theoretical guarantees in Section 2.3 to address potential concerns regarding the proposed framework. Specifically, they show that symmetrizing the denoiser does not increase the DDPM training risk, and that symmetry-respecting averaging preserves the false positive rate without degrading the detection trade-off.
- Clarity and coherence: The paper clearly presents its core objective and maintains a focused narrative throughout. Most components of the proposed framework appear well-motivated (e. g. , the introduction of equivariant noise coupling to enforce symmetry in the diffusion process). The experimental results provide convincing empirical support for the claims. In addition, the inclusion of pseudo code (Algorithm 1) improves the reproducibility of the method.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The overall methodological novelty appears somewhat limited, as most components are adaptations of existing techniques. The main contribution lies primarily in their integration rather than the introduction of fundamentally new methodological ideas.
- The theoretical guarantees rely on strong symmetry assumptions that may not strictly hold for real cytology data, where cell morphology and acquisition conditions are not perfectly D4-invariant.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Overall, the paper is well structured and clearly written. I particularly appreciated that many of the questions that arise while reading are addressed within the paper itself. One minor suggestion is that the framework illustration in Fig. 1 could be made more detailed, which would help readers better understand the overall pipeline at an early stage of the paper.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed method presents a well-motivated problem and a corresponding solution that fits well within the MIC category. Specifically, the paper addresses the problem of enforcing equivariance in diffusion-based anomaly detection, which is an interesting and relevant direction. Although the solution may appear relatively simple, it is clearly formulated and well justified. Furthermore, the authors provide both solid theoretical explanations and extensive experimental results, which together strongly support the effectiveness of the proposed approach.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Although I raised concerns regarding limited methodological novelty and the assumptions related to D4-invariance, I believe that the paper’s well-motivated problem definition and the academic value of the proposed framework outweigh these weakness.
Review #2
- Please describe the contribution of the paper
This paper proposes a D4-equivariant diffusion framework for anomaly detection in computational cytology. Since in cytology the signal is cell-centric, standard diffusion models may produce transformation dependent reconstructions and unstable anomaly scores. To address this, the paper introduce a dihedral-equivariant diffusion framework that enforces D4 symmetry throughout partial-diffusion reconstruction via a U-Net denoiser with D4-equivariant architecture, together with inference time equivariant noise coupling (EN) and frame averaging (FA). Experiments on two public cytology datasets achieve higher AUC and retrieve more abnormal cells in the top K predictions.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The observation that diagnostic class should not change under rotations and flips for single-cell patches is reasonable, while standard generative detectors treat rotated views as distinct inputs, inducing orientation-dependent score variation. The paper is well motivated from a computational cytology perspective. 2.The method is clearly presented and the algorithmic is easy to follow. 3.The ablation study includes the comparison between FA-only, FA+EN, together with the reported reduction in anomaly-score variance across rotations and flips, which provides evidence that the proposed method improve transformation consistency.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The novelty is limited. Based on Section 2, the main components are U-Net denoiser with a D4-equivariant architecture [9,16], diffusion anomaly detection via partial reconstruction [22,35], and inference-time symmetry enforcement through equivariant noise coupling (EN) and frame averaging (FA) [22,35], all of them are individually known. Consequently, the proposed approach seems more like a combination of known components than a substantially new diffusion framework.
2.The experiment results are inconsistent across two datasets and often small against the strongest baselines, which weakens the central claim and raises concerns about the generalization of the proposed method.
The paper on Page 8 shows that the D4 model is statistically inconclusive on AML LMU. So the positive result is mainly on MLL. Does this suggest that the effectiveness of the method may depend on the specific dataset rather than reflecting a robust improvement? As a result, it remains unclear whether the proposed method generalizes reliably across different cytology dataset, or whether its effectiveness is sensitive to the underlying data distribution.
3.There are some inconsistent ablation study results across two datasets. Section 3.4 states FA-only slightly degrades AUC on MLL, while FA+EN improves it, why FA+EN performs worse than FA-only in terms of AUC on AML LMU? Besides, the D4 convolutional U-Net slightly outperforms the attention-augmented variant in MLL but not on AML LMU. This suggests that the performance gains may be sensitive to implementation details and dataset properties .
4.The theoretical section is supportive, but some claims are insufficiently justified and therefore not fully convincing. Section 2.3 gives intuitive statements about D_4 averaging can not increase DDPM risk and invariant detection under class-conditional symmetry, but no proof is provided in the paper. Besides, the analysis relies on strong symmetry assumptions that may not hold in real cytology patches.
5.The qualitative results in Figure 2 raise concerns about the practical localization ability of the method. The displayed MSE values are very small and the anomaly maps appear weak and poorly localized. This suggests that the method may be capturing patch-level differences without producing sufficiently clear or clinically meaningful pixel-level abnormality maps.
6.In section 2.1, the notation is somewhat inconsistent: x_0 is introduced as coming from the normal training distribution , but is also used to denote a test sample.
7.Some typos: in section 2.3, “Invariant detection under class-conditional symmetry:Let Y…”
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.The novelty is limited. Based on Section 2, the main components are U-Net denoiser with a D4-equivariant architecture [9,16], diffusion anomaly detection via partial reconstruction [22,35], and inference-time symmetry enforcement through equivariant noise coupling (EN) and frame averaging (FA) [22,35], all of them are individually known. Consequently, the proposed approach seems more like a combination of known components than a substantially new diffusion framework.
2.The experiment results are inconsistent across two datasets and often small against the strongest baselines, which weakens the central claim and raises concerns about the generalization of the proposed method.
The paper on Page 8 shows that the D4 model is statistically inconclusive on AML LMU. So the positive result is mainly on MLL. Does this suggest that the effectiveness of the method may depend on the specific dataset rather than reflecting a robust improvement? As a result, it remains unclear whether the proposed method generalizes reliably across different cytology dataset, or whether its effectiveness is sensitive to the underlying data distribution.
3.There are some inconsistent ablation study results across two datasets. Section 3.4 states FA-only slightly degrades AUC on MLL, while FA+EN improves it, why FA+EN performs worse than FA-only in terms of AUC on AML LMU? Besides, the D4 convolutional U-Net slightly outperforms the attention-augmented variant in MLL but not on AML LMU. This suggests that the performance gains may be sensitive to implementation details and dataset properties .
4.The theoretical section is supportive, but some claims are insufficiently justified and therefore not fully convincing. Section 2.3 gives intuitive statements about D_4 averaging can not increase DDPM risk and invariant detection under class-conditional symmetry, but no proof is provided in the paper. Besides, the analysis relies on strong symmetry assumptions that may not hold in real cytology patches.
5.The qualitative results in Figure 2 raise concerns about the practical localization ability of the method. The displayed MSE values are very small and the anomaly maps appear weak and poorly localized. This suggests that the method may be capturing patch-level differences without producing sufficiently clear or clinically meaningful pixel-level abnormality maps.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The following issues remain material and prevent me from raising my score. 1.Generalization claim is not supported on AML-LMU. Reframing the endpoint as TP@K is reasonable, but the underlying fact is unchanged: on AML-LMU, as the authors themselves acknowledge, the results are “statistically inconclusive”, which suggests that the performance gains may be sensitive to implementation details and dataset properties. These results do not justify the claim of a robust, generalizable improvement. 2.Theoretical results remain unproven. The sketch that Jensen’s inequality applies to the group-averaged denoiser, and that the detection result follows from class-conditional label symmetry, is helpful as intuition, but the rebuttal does not actually contain the derivation. Combined with the limited methodological novelty, I do not intend to raise my score.
Review #3
- Please describe the contribution of the paper
This paper proposes a D4-equivariant diffusion framework for anomaly detection in computational cytology, targeting the key issue that cell-level morphology is invariant to rotations and reflections, while standard diffusion models are not. The method introduces: 1.D4-equivariant U-Net architectures to enforce symmetry at the model level. 2.Equivariant Noise (EN) and Frame Averaging (FA) at inference to ensure transformation-consistent sampling trajectories. 3.A partial diffusion reconstruction framework (AnoDDPM-style) to compute anomaly scores.
Results show consistent improvements over multiple baselines, including diffusion and GAN-based anomaly detectors.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The application of dihedral group equivariance (D4) to diffusion models specifically to address the biological invariance of single-cell cytology is well-motivated.
- The authors evaluate the method using a highly realistic metric for budgeted clinical review workflows: the number of true positives in the top-K retrieved anomalous predictions (e.g., TP100/200/400).
- Section 2.3 provides risk bound via Jensen’s inequality and justification of invariant scoring under group symmetry. This significantly strengthens the contribution beyond empirical work.
- The study benchmarks against a wide array of relevant methods, including multiple instance learning (ItS2CLR), deep one-class models (Deep SVDD), and several generative detectors like f-AnoGAN, THOR, and BerDiff. AUC improves from 0.63 → 0.68, TP400 improves from 57 → 78 (large practical gain).
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- While the performance gains on the MLL dataset are clinically and statistically significant, the improvements on the AML LMU dataset are statistically inconclusive. This limits the strength of the generalizability claims across different cytology modalities.
- The ablation studies reveal that the $D_{4}$ convolutional U-Net without attention performs best on the MLL dataset, whereas the attention-augmented variant achieves the strongest performance on the AML LMU dataset. The paper lacks a deeper investigation or hypothesis regarding why this architectural discrepancy exists between the two datasets.
- Only two cytology datasets, where both are similar modalities (single-cell patches). Missing whole-slide validation and cross-domain generalization
- Assumes perfect D4 invariance, which may not fully hold for staining artifacts or slide preparation biases, Need analysis of symmetry violations.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the methodological novelty is somewhat incremental, the combination of theory, domain insight, and practical impact makes this work above the acceptance threshold.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and AC for recognizing the motivation of our work, its clear formulation, theoretical grounding, strong evaluation, top-K gains, and practical relevance.
A. Novelty (R2,R3,R4): Our contribution is the first systematic study of dihedral equivariance for diffusion-based unsupervised anomaly detection in cytology, with a unified algorithm treating architectural and inference-time equivariance as complementary routes to the same invariance property. This is a non-trivial adaptation to partial-diffusion anomaly scoring, where the goal is not generation/translation but stable top-K ranking under budgeted cytologist review. Prior cytology or diffusion-based medical anomaly-detection work has not enforced D4 structure or measured score variance under symmetry. We agree that equivariant convolutions, partial diffusion, and test-time averaging have precedents; our novelty lies in their task-specific integration and validation for cytology anomaly detection. B. Statistical significance (R3,R4): Our primary clinical endpoint is budgeted top-K retrieval: reducing thousands of candidate cells to a few hundred suspicious patches for cytologist review. AUC summarizes global ranking, while TP@K measures this operational goal. On MLL, the gains are clear: FA+EN improves AnoDDPM from TP400/200/100 = 57/30/18 to 78/53/34, with significant AUC and TP400 gains (p<0.001 and p=0.006). On AML-LMU, the margin is statistically inconclusive (AUC p≈0.125, TP400 p≈0.275), but D4 variants still give the best or tied-best top-K results: TP400=325, TP200=162, TP100=86.Thus, AML-LMU shows favorable but not statistically conclusive gains, not a contradiction of the method. C. Architectural ablation differences (R3,R4): Our claim is not that one D4 backbone is universally optimal, but that D4 structure improves transformation-consistent scoring. The attention variant mainly adds capacity/context modeling. Thus, D4-conv may suffice for MLL’s smaller 250x250 bone-marrow patches, while attention can help on AML-LMU’s larger 400x400 peripheral-blood patches with greater visual diversity. Both support the trend: D4-conv improves MLL over AnoDDPM (AUC/TP400 0.675/74 vs 0.633/57), and D4-conv+attn improves AML-LMU (0.720/325 vs 0.694/311). We clarify that attention is dataset-dependent, not the core contribution. D. D4 assumptions and proof clarification (R2,R3,R4): We agree real cytology data are not perfectly D4-invariant due to staining/acquisition effects. Our assumption is instead class invariance: rotating/flipping a centered square single-cell patch should not change its normal/abnormal label. Sec. 2.3 is an idealized justification for symmetry-respecting scoring; the empirical question is whether this reduces orientation-induced score variation, which our D4 variants do while improving top-K retrieval. For R3, the DDPM-risk result follows from Jensen’s inequality applied to the group-averaged denoiser, and the detection result from class-conditional label symmetry. We will clarify the assumption in the proof sketch. E. Qualitative anomaly maps (R3): Fig. 2 is only a qualitative visualization; the method and metrics use scalar patch-level anomaly scores. Unlike MRI, OCT or X-ray lesions, cytology abnormalities are often not sharply localized, but expressed through nuclear size/shape, chromatin texture, nuclear-cytoplasmic ratio, cytoplasm, staining, and overall atypia. Thus, maps show reconstruction-deviation patterns within suspicious cells, not pixel-level segmentation. Since our goal is to reduce thousands of patches to a few hundred for cytologist review, patch-level top-K retrieval is the endpoint. We have revised the caption accordingly. We will address other comments (x0 notation, Sec. 2.3 typos, Fig. 1/2 captions). For R4’s whole-slide/cross-domain point, we agree this is important future work; this study focuses on controlled single-cell patch evaluation for rare-cell retrieval before slide-level deployment.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Two reviewers gave rather positive evaluations of the paper with high confidence. However, regarding the negtive evaluation and other mentioned issues, the authors should should meticulously address reviewers’ concerns. Especially, the concerns about the limited novelty, inconsistent data reported and insufficient justification. Other concerns about experiment result analysis are also needed to response.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Reviewers noted clear motivation and generally positive evaluations. The rebuttal addressed several clarifications and mitigated part of these issues. Overall, though several concerns remain, the rebuttal supports the final acceptance with two rather positive recommendation.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This work develops a method for anomaly detection in cytology images using a diffusion model. In particular, the approach develops a D_4 equivariant diffusion model to ensure that the quasi-normal reconstructions which are used for abnormality assessment do not change based on rotations by 90, 180, 270 degrees or flips.
The approach is tested on two different datasets with statistically significant improvements with respect to competing methods on one of the two datasets.
In general, the reviewers (and this meta reviewer) appreciated the goal of using this approach for more efficient screening of cytological abnormalities by experts and the combination of theory, domain insight, and practical utility.
While concerns were raised (and somewhat remained after the rebuttal) with respect to novelty and the statistically inconclusive results for one of the experimental datasets, the approach itself is sufficiently interesting in my opinion to be presented at MICCAI.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The rebuttal doesn’t fully address reviewers’ concerns.
