List of Papers Browse by Subject Areas Author List
Abstract
Unsupervised anomaly detection (UAD) aims to localize abnormal regions in medical scans without pixel-level annotations. A typical strategy seeks to reconstruct a pseudo-healthy image that preserves subject-specific anatomy. Recently, diffusion models have been proposed to perform UAD. However, these methods rely on heuristic noise schedules or synthetic corruptions to balance subject-specificity and anomaly removal.
In this work, we propose an alternative formulation of UAD as a Bayesian inverse problem under a diffusion prior. First, we introduce a latent spatial anomaly mask that models pixel-wise consistency between a test image and its latent corresponding pseudo-healthy image. Then, we propose an approximation of the unknown generation process that links healthy anatomy, anomalies, and the observed image, enabling a well-defined likelihood within the Bayesian framework. Building on recent advances in diffusion-based inverse problem methods, we jointly infer the pseudo-healthy image and the anomaly mask via annealed posterior sampling. We evaluate our approach on FDG PET (ADNI) and FLAIR MRI (BraTS 2021), demonstrating improved anomaly localization performance compared to other diffusion-based approaches and validating the contribution of our introduced model. Our code is available at https://github.com/HuguesRoy/UAD_DAPS
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4982_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HuguesRoy/UAD_DAPS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{RoyHug_Unsupervised_MICCAI2026,
author = { Roy, Hugues AND Dorent, Reuben AND Burgos, Ninon},
title = { { Unsupervised Brain Anomaly Detection as a Bayesian Inverse Problem with Diffusion Prior } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes an inverse-problem formulation of anomaly detection, where a diffusion model is used to describe the prior distribution over images and the latent clean image and anomaly mask are inferred via an annealed posterior sampling algorithm. This is demonstrated in 2D PET and MRI benchmarks and shows competitive performance.
- 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.
A novel probabilistic algorithm for mask inference is introduced, which seems mathematically principled and well justified. The procedure is described clearly with a good level of mathematical detail. The results indicate a good level of performance, both quantitatively and qualitatively and is competitive with other works.
- 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.
It’s a bit unclear why only the best possible Dice is reported? (with standard deviation across subjects/slices?) without including mean, and std-dev across samples? In practice we don’t have access to an oracle to select so it would be good to understand the average, and worst case scenario. It would also be instructive to see the performance on some anomaly free slices. Some information about the diversity of inferred masks would be interesting.
[1] should also be referenced and compared against.
[1] Beizaee, Farzad, et al. “Mad-ad: Masked diffusion for unsupervised brain anomaly detection. “ International Conference on Information Processing in Medical Imaging. Cham: Springer Nature Switzerland, 2025.
- 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
Page 6: [Dice] rather than a citation.
It would be good to understand whether the DAPS algorithm was a required component, or if the mask inference procedure is partially agnostic to the form of diffusion based inverse problem method.
In the graphical model in Fig 1, it would be beneficial to denote y in a different format to make it obvious that it’s the only variable that is observed, and everything else is latent.
- 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 algorithm is novel (to the best of my understanding), well principled and achieves good performance. The paper is also clearly presented. I have some reservations about only reporting the max Dice, but this should be straightforward to resolve, and there is a paper to cite.
- 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
Review #2
- Please describe the contribution of the paper
This work’s main contribution is to recast unsupervised brain anomaly detection as explicit Bayesian posterior inference under a diffusion prior. In particular, they introduced a latent anomaly mask and an approximate likelihood using them that links the image to a pseudo-healthy image. The samples of the pseudo heathy image, mask, and the anomaly are derived using Langevin dynamics using DAPS.
- 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.
- Current diffusion-based UAD methods methods (e.g. THOR, AutoDDPM) also use anomaly maps derived during inference diffusion inference. The main distinction of this work is to pose the recovery of the mask and pseudo-healthy image as a single Bayesian inference rather than a heuristic combinations of mask and diffusion inference. This is a much more elegant approach, and it translates to improved performance as shown by their results.
- The qualitative results in Fig. 2 are impressive and clearly demonstrate the benefits of the proposed 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.
1.The paper summarizes current diffusion-based UAD methods in one sentence without providing any details on their mechanisms and the distinction with their formulation. I think describing how they work, and how they differ from the proposed scheme would help the readers appreciate the methodological contributions.
2.The focus is on having a single elegant formulation, which is a clear strength. However, the authors make several heuristic choices, which are not well motivated. In particular, the authors vary the weightings and contributions of different terms arbitrarily during the annealing process, which are not theoretically justified. For example
a. Why do they choose the likelihood variance to vary depending on the scale (k). I understand the prior variance could vary depending on the scale, but it is not common to vary the likelihood variance. Why would this be justified, and what solution will the algorithm converge to.
b. They mention that the prior weight is tuned depending on the scale. They should make the exact dependence more explicit. This again seems to be a deviation from the “true Bayesian formulation”.
c. The step sizes in (6) and (7) are selected at each step based on gradient magnitudes. Please state the exact formula used for the step sizes.
3.It would have been helpful to include a figure showing the evolution of the mask and the pseudo-healthy image as a function of annealing steps.
- 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.
(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 paper builds upon the existing diffusion approaches for anomaly detection by introducing a combined Bayesian formulation that is less heuristic. Their results clearly show the benefit of the combined formulation. While they also make some heuristic choices in the implementation, I think the overall contribution merits acceptance.
- 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
Review #3
- Please describe the contribution of the paper
This paper proposes a novel formulation of unsupervised anomaly detection as a Bayesian inverse problem with a diffusion prior. A latent spatial anomaly mask is introduced to model pixel-wise consistency between the observed image and its pseudo-healthy reconstruction, and the unknown forward model is approximated via a mask-weighted likelihood. Joint inference of the healthy image and anomaly mask is performed via annealed posterior sampling with Langevin dynamics, building on the DAPS framework. The method is evaluated on brain MRI (BraTS 2021) and FDG PET (ADNI) and compared against a range of reconstruction-based and diffusion-based UAD 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.
-
The Bayesian inverse problem formulation is principled and well-motivated, addressing a genuine limitation of existing diffusion-based UAD methods. The theoretical grounding is clearly presented.
-
The posterior factorisation cleanly disentangles anomaly information from healthy anatomy by separately modelling a healthy image prior and an anomaly mask, which is a conceptually appealing and novel contribution.
-
- 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.
-
Small anomalies are entirely excluded from the BraTS evaluation, with only medium and large lesions considered. This is a significant omission, as detection of subtle anomalies is arguably the most clinically relevant and challenging regime, and precisely the setting where the generative approach and global mask modelling are most likely to struggle. Without this evaluation, the practical utility of the method remains unclear.
-
BraTS T1 and T2 modalities are not included, limiting the evaluation to FLAIR only and neglecting low-contrast anomaly settings where the method may behave differently.
-
The approximate forward model assumes that anomalies correspond to high-variance Gaussian deviations from healthy samples (Eq. 3), which is generally not true in practice where anomalies are structured. While acknowledged in the conclusion, this is a fundamental modelling limitation that deserves more prominent discussion and, if possible, empirical analysis of its impact.
-
The independence assumption between xt and m in the intermediate posteriors (Eq. 2) is not justified. Unlike x0, xt may carry anomaly localisation information, and assuming independence here could introduce a non-trivial approximation error that is not discussed.
-
Slice-level anomaly scoring is not discussed and is omitted from the evaluation entirely, despite being a clinically relevant and commonly reported metric in the UAD literature.
-
The baseline comparison is limited to diffusion-based UAD methods and a small number of reconstruction-based approaches. Established feature-distribution-based (self-supervised) methods such as PaDiM [1], PatchCore [2] or PII/NSA [3-4], which have shown strong performance in anomaly detection tasks, are not included. Without these comparisons it is difficult to assess where the proposed method truly stands relative to the broader UAD landscape.
-
The method involves numerous hyperparameters (λ0, λc, μ, c, step sizes ηx and ηa, number of Langevin iterations, number of intermediate posterior steps), and no sensitivity analysis is provided. The description of how ηx and ηa are updated is particularly limited, raising reproducibility concerns.
-
The ablation study is conducted only on BraTS large lesions, leaving open whether the relative contributions of each component hold across lesion sizes, modalities, and the excluded small anomaly regime.
-
The PET anomalies are entirely synthetic, which limits the clinical validity of that part of the evaluation. This is not sufficiently discussed.
-
The method is evaluated on 2D slices only, and the extension to 3D is deferred to future work, which is a notable practical limitation given that most clinical use cases involve volumetric data.
[1] Defard, T., Setkov, A., Loesch, A., & Audigier, R. (2021, January). Padim: a patch distribution modeling framework for anomaly detection and localization. In International conference on pattern recognition (pp. 475-489). Cham: Springer International Publishing.
[2] Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., & Gehler, P. (2022). Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 14318-14328).
[3] Tan, J., Hou, B., Day, T., Simpson, J., Rueckert, D., & Kainz, B. (2021, September). Detecting outliers with poisson image interpolation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 581-591). Cham: Springer International Publishing.
[4] Schlüter, H. M., Tan, J., Hou, B., & Kainz, B. (2022, October). Natural synthetic anomalies for self-supervised anomaly detection and localization. In European Conference on Computer Vision (pp. 474-489). Cham: Springer Nature Switzerland.
-
- 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
-
Including small anomalies in the BraTS evaluation is strongly encouraged, as this is the regime most relevant to assessing the practical limitations of the approach.
-
Extending the baseline comparison to include feature-distribution-based methods would provide a more complete and informative evaluation of the proposed approach relative to the broader state of the art.
-
Extending the evaluation to additional BraTS modalities (T1, T2) would provide a more complete picture of the method’s robustness to anomaly contrast.
-
A more thorough discussion and ideally empirical analysis of the Gaussian anomaly assumption and the independence assumption in Eq. 2 would strengthen the theoretical transparency of the paper.
-
A sensitivity analysis of the number of intermediate sampling steps and Langevin iterations, given their linear impact on inference time, would be informative and practically useful. Including slice-level scoring would bring the evaluation in line with standard practice in the UAD literature.
-
An interesting direction worth exploring would be to use the posterior probability of an approximately empty mask (m ≈ 1) as a normality score, which flows naturally from the proposed framework and could provide a principled image-level detection metric.
-
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The methodology is novel, rigorously derived, and the probabilistic perspective on UAD is a genuinely valuable contribution that could inspire future work. However, the evaluation has crucial gaps, most notably the exclusion of small anomalies, slice-level scoring and only sparse comparison to SOTA UAD methods, that prevent a confident assessment of the method’s practical utility. These gaps are addressable without fundamental changes to the method.
- 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 for their constructive feedback. Since the paper has been provisionally accepted, we will use the revision to clarify the following modeling choices and limitations.
[R4] Independence between $x_{t_k}$ and $m$. We agree that $x_{t_k}$ may retain residual information about anomalies. However, following DAPS, the noisy diffusion state $x_{t_k}$ is only introduced as an auxiliary variable to approximate the intractable image prior $p(x_0)$ via the conditional diffusion prior $p_k(x_0 \mid x_{t_k})$. We additionally assume independence from $m$ as a tractable approximation. The empirical performance across both datasets suggests this approximation does not critically impair the method in practice. We have made this approximation explicit in the revised manuscript.
[R4] Approximate the forward model. We would like to clarify the distinct roles of the two components of our forward model. While Eq. 3 utilizes an isotropic Gaussian likelihood, this is only one component of our model. The spatial structure of anomalies is explicitly captured by the random Gaussian field prior on the mask $m$. As shown by the ablation without spatial regularization in the mask prior (Table 3-(C)), the isotropic Gaussian likelihood alone struggles to capture structured anomalies, highlighting the importance of the structured prior.
[R3] Annealed likelihood variance. Unlike standard inverse problems with a known forward model, both the reconstruction and the anomaly mask are uncertain in our setting. Inspired by DAPS, we use a sequence of intermediate likelihoods whose variance evolves jointly with the conditional diffusion prior across scales. A larger variance in early iterations allows broader exploration of reconstruction-mask pairs, while a smaller variance later enforces closer consistency with the measurement. The algorithm targets the final annealed approximate posterior associated with the last variance level. The fixed-variance ablation in Table 3-(B) supports the usefulness of this annealing.
[R3] Scale-dependent $\lambda_0$. Since the likelihood term scales as $1/r_k^2$, keeping $\lambda_0$ fixed would change the relative strength of the likelihood and mask regularization during annealing. We adapt $\lambda_0(k)$ across scales to maintain a consistent balance between these terms in the intermediate posteriors. This is part of the annealed approximation used for sampling, rather than a claim that the underlying Bayesian prior itself is scale-dependent.
[R3-R4] Langevin step sizes. As in DAPS, the effective step size decreases as the intermediate posteriors become sharper. Since the energy gradients with respect to the healthy image $x_0$ and the mask unconstrained logits $a$ can have different scales, we adapt the step sizes $\eta_x$ and $\eta_a$ inversely to the gradient of the energies $|\nabla x_0 U_k|2^2$ and $|\nabla a U_k|_2^2$, respectively, with clipping for numerical stability.
[R4] PET evaluation and additional MRI sequences. We agree that synthetic PET anomalies do not capture the full clinical variability of real disease. However, our goal was to complement the BraTS FLAIR tumor experiments with a different modality and disease pattern: FDG PET with Alzheimer-like hypometabolism. The synthetic setting provides ground truth and enables quantitative comparison, while still revealing differences between methods.
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 were strongly positive. R4 is weakly negative, but raises issues that may be addressed. However, I also agree that R4’s raised issues limit scope and impact.
