Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Anomaly detection in medical images is useful for early diagnosis and treatment planning. Unsupervised methods are attractive because they avoid manual annotations and generalize well to unseen anomalies. However, existing unsupervised methods face fundamental limitations: reconstruction-based methods often suffer from pixel-level mismatch of generated pseudo-normal images, and embedding-based methods often fail to separate normal and abnormal images sufficiently in embedding spaces. In this paper, we propose an unsupervised method named InvDetect to address both limitations. InvDetect performs anomaly detection in a noise latent space induced by Denoising Diffusion Implicit Model (DDIM) inversion, which avoids generating pseudo-normal images thus prevents pixel-level mismatch. By training DDIM exclusively on normal image patches, InvDetect constructs a structured noise latent space in which normal patches form a compact cluster and abnormal patches deviate significantly. This yields better separation between normal and abnormal patches to produce a more reliable pixel-wise anomaly map. At last, we refine the anomaly map by enforcing spatial contiguity of abnormal regions, which reduces isolated false positives and false negatives in the final detection result. Experiments on four real-world medical imaging datasets demonstrate that InvDetect consistently outperforms eighteen prior unsupervised anomaly detection methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0456_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lele0007/InvDetect

Link to the Dataset(s)

BraTS2021: https://www.cancerimagingarchive.net/analysis-result/rsna-asnr-miccai-brats-2021/ BTCV: https://www.synapse.org/Synapse:syn3193805/wiki/217789 RESC: https://github.com/AIChallenger/AI_Challenger_2018 IDRiD: https://ieee-dataport.org/open-access/indian-diabetic-retinopathy-image-dataset-idrid

BibTex

@InProceedings{MaXin_InvDetect_MICCAI2026,
        author = { Ma, Xinyu AND Ma, Jinhui AND He, Shiqi AND Pei, Jian AND Chu, Lingyang},
        title = { { InvDetect: Unsupervised Medical Anomaly Detection in the Noise Latent Space of DDIM } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution of the paper is a new formulation of unsupervised medical anomaly detection in the noise latent space induced by DDIM inversion, designed to address two major limitations of prior work at once: pixel-level mismatch in reconstruction-based methods and insufficient normal/abnormal separation in embedding-based methods. The authors train DDIM only on normal patches, invert unseen patches into the noise latent space, score them with a lightweight one-class SVM, and then refine the resulting anomaly map using a latent-similarity-based spatial contiguity refinement (SCR) module. This yields a conceptually clean pipeline that avoids pseudo-normal image synthesis while still producing pixel-wise anomaly localization.

  • 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.

    First, the paper is motivated by a very clear problem formulation. It identifies two concrete failure modes in prior unsupervised anomaly detection pseudo-normal reconstruction mismatch and weak abnormality separation in generic embedding spaces and proposes a method that directly targets both. The idea of avoiding pseudo-normal image generation altogether and instead detecting anomalies directly in an inversion-induced latent space is simple yet compelling.

    Second, the empirical evaluation is broad. The method is tested across four very different medical datasets/modalities and compared against 18 prior baselines, including both reconstruction-based and embedding-based approaches. The ablation comparing InvDetect and InvDetect w/o SCR also supports the importance of the proposed refinement step.

    Third, the paper includes a useful runtime-versus-Dice analysis rather than reporting accuracy alone. This helps position the method in terms of practical efficiency and shows that the approach remains competitive among stronger but slower methods.

  • 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 main weakness of the paper is that its central hypothesis is not validated directly enough. The claim that normal and abnormal patches become better separated in the DDIM noise latent space is intuitive and plausible, but the evidence remains largely indirect, relying mostly on final detection performance rather than explicit latent-space analysis. The paper would be much stronger with more direct support, such as latent distribution visualizations, separation statistics, or calibration-style analyses.

    In addition, several components of the pipeline are not novel in isolation: one-class SVM is a classical normality estimator, and patch-based aggregation is also standard, so the main novelty appears to lie primarily in the inversion-induced noise latent space itself. This makes stronger component-level ablations particularly important; comparisons against alternative scoring models or normality estimators would help clarify which parts of the method are truly essential.

    The method also depends heavily on a fixed patch-based design, yet no sensitivity analysis is provided for patch size or stride.

    Finally, reproducibility and practical applicability would be improved by more detailed reporting of the DDIM inversion settings and clearer analysis of computational scaling, especially since runtime is dominated by DDIM inversion and grows with the number of patches.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Overall, I found this to be a solid and well-motivated paper with strong empirical performance. The idea of avoiding pseudo-normal image generation and instead performing anomaly scoring directly in an inversion-induced noise latent space is both interesting and practically meaningful. The multi-dataset evaluation and the SCR ablation are also convincing strengths. My main reservations are that the central latent-separation claim is not directly analyzed in enough depth, and some additional ablations would strengthen the paper. For these reasons, I place the paper slightly above the acceptance threshold as a Weak Accept.

  • 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



Review #2

  • Please describe the contribution of the paper

    This paper proposes a novel approach to use the latent space of an inverted DDIM for anomaly detection. Concretely, after training the DDIM on patches extracted from images of healthy patients, the latents corresponding to normal patches should form a Gaussian cluster, to which a one class SVM is fitted to model the normal distribution. Additionally, the authors propose to binarize the anomaly score map using an energy based formulation that aligns the segmentation mask with the anomaly scores while including neighborhood consistency.

  • 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.
    • Although the proposed method is in the mid range in terms of runtime, it runs much faster than the image reconstruction methods with newer generative models and it performs consistently well over all evaluated domains in terms of anomaly segmentation scores.
    • The spatial contiguity refinement (SCR) is a novel way to threshold the anomaly score map in an unsupervised manner and is a good contribution to the field.
    • The manuscript empirically demonstrates that the latent space of DDIM is useful for separating between normal and abnormal samples and intuitively explain this with the way the latent space is constructed by DDIM inversion.
  • 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.
    • Most state of the art methods evaluate the anomaly score maps without thresholding with metrics such as AUROC or AUPR. Reported anomaly segmentation scores are usually obtained by using a fixed threshold, a threshold such that the F1 score on a small labeled set is maximized, or a threshold to reach a predefined false positive rate on a small set of normal images. The manuscript provides only segmentation scores and does not explain how the threshold is set for the state of the art methods and for the proposed method without SCR. Therefore, it is impossible to evaluate if the comparison is fair.
    • Both Dice score and precision are highly affected by class imbalance, which is usually the case in anomaly detection.
    • The manuscript does not justify why the additional training of a one class SVM is required instead of using the probability density of the multivariate normal distribution, which should be the distribution of the latents corresponding to normal patches.
    • It was not evaluated if SCR preserves small anomalies or if they are removed. From the formulation, it is more likely that small anomalies are removed or suppressed, which would also be in line with the smaller performance improvement on the BTCV dataset compared to the proposed method without SCR, as BTCV tends to contain smaller anomalies.
    • Using only abnormal images in the test set is quite atypical for anomaly detection, which could be beneficial for methods that process the image in a patch-wise fashion as normal regions are much less affected by abnormal regions in the same image. While this is a desirable property, its impact on the scores might be over-amplified in comparison to the typical setup with normal + abnormal samples during testing.
  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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
    • Please add references to the literature for DDIM and particularly DDIM inversion, especially as there is almost no explanation provided.
    • The statement that the proposed method outperforms embedding-based methods because the latter do not explicitly separate normal and abnormal embeddings is vague, because the proposed method does not separate the two classes explicitly either.
    • PaDiM would most likely be considered as a method using a probabilistic model instead of a memory bank in the introduction.
    • License of the datasets should rather be a footnote than a reference.
  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Despite the above mentioned issues in the evaluation and comparison, the usefulness the latent space of DDIMs for anomaly detection and the unsupervised segmentation with SCR are of great interest to the anomaly detection communities in MICCAI and possibly beyond the medical domain.

  • 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



Review #3

  • Please describe the contribution of the paper

    The paper proposes InvDetect, an unsupervised medical anomaly detection method that operates in the DDIM inversion latent space instead of reconstructing pseudo-normal images. A DDIM is trained on normal patches, a one-class SVM scores inverted latents, and a graph-cut spatial refinement yields the final mask. Experiments cover four medical datasets against a broad set of unsupervised baselines.

  • 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.
    • Doing UAD in the DDIM inversion latent space is a clean idea that sidesteps the pixel-mismatch issue of reconstruction-based methods.
    • The pipeline is simple and lightweight (one-class SVM + graph-cut) with practical runtime.
    • Evaluation spans four modalities and a reasonably broad set of baselines.
  • 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 central claim that normal latents concentrate near N(0, I) while abnormal ones drift away is taken largely on faith. Some direct evidence (norm distributions, t-SNE/UMAP, or a latent-space AUROC) would make the motivation much more convincing.
    • The evaluation leans on Dice/Precision from thresholded binary masks, which are sensitive to post-processing. Since the proposed method has a dedicated graph-cut step and most baselines do not, the comparison may be unfair. Is there any other metric can be used without any post-processing?
    • Dataset curation is underspecified. BraTS2021 is typically tumor-only so the source of “normal” training data is unclear, and for BTCV the anomaly definition is not obvious — both matter a lot in a normal-only setup.
    • The choice of a one-class SVM in a high-dimensional latent space is not compared against simpler density estimators (e.g., Mahalanobis), and hyperparameter handling for SCR reads inconsistently between per-dataset tuning and fixed values.
    • The method is purely 2D, which is a real limitation for volumetric CT/MRI where 3D context would likely help both separability and spatial contiguity. Key implementation details (DDIM steps, noise schedule, patch size/stride sensitivity) are also unclear.
    • Minor: there is no such thing as ‘train a DDIM’, which is a sampling method applied on a trained denoiser. The authors are recommended to correct the phrasing.
  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The idea is nice and the results look promising, but the experimental rigor is not quite there: the central latent-space claim lacks direct evidence, the evaluation protocol can bias comparisons, and dataset definitions need tightening. These are fixable, so I lean towards weak accept.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 all the reviewers for their valuable comments. We will carefully address the raised comments in the camera-ready version, following the camera-ready submission guidelines.



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.

    All three reviewers recommend acceptance. The core idea of performing UAD directly in the DDIM inversion noise latent space, bypassing pseudo-normal image synthesis, is clean and well-motivated, and the SCR module is a genuine contribution to unsupervised anomaly localization. Reviewers share one common concern worth addressing in the final version: the central claim that normal latents concentrate near N(0,I) while abnormal ones drift away lacks direct empirical support. Please revise these claims. Latent distribution visualizations or a latent-space AUROC would substantially strengthen the paper.



back to top