List of Papers Browse by Subject Areas Author List
Abstract
Diffusion MRI (dMRI) is an advanced imaging technique that can characterize tissue microstructure and map white matter connections in vivo. Anomaly detection of brain lesions, such as tumors, white matter hyperintensities, and intracerebral hemorrhage, is a critical task in dMRI, supporting neurological disease diagnosis and surgical planning. Unsupervised anomaly detection leveraging uncertainty models offers an effective approach for distinguishing diffusion measures of abnormal tissues from those of normal structures. However, this task is complicated by acquisition noise and cross-protocol variability, which can obscure subtle pathological signals. In this study, we propose DDecomp, a masked-autoencoder-based framework trained on healthy data that decomposes predictive uncertainty and leverages the epistemic component as an out-of-distribution cue for unsupervised anomaly detection. The key idea is to disentangle uncertainty arising from inherent data noise from uncertainty indicative of model mismatch to pathology, thereby generating an epistemic map that is specific to the anomaly. Once trained, the model generalizes to detect multiple lesion types within a unified framework. In the experiments, we evaluate our method on three diverse clinical cohorts (tumors, white matter hyperintensities, and intracerebral hemorrhage) acquired with different scanners and protocols. Across all cohorts, our approach achieves robust lesion detection and consistently outperforms state-of-the-art baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6538_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/WHOOOOOS/DDecomp
Link to the Dataset(s)
N/A
BibTex
@InProceedings{FuZhi_DDecomp_MICCAI2026,
author = { Fu, ZhiJia AND Zhang, Wei AND Zheng, RuiXi AND Li, YiJie AND Huang, GuanLin AND Ning, Lipeng AND O’Donnell, Lauren J. AND Pasternak, Ofer AND Zhang, Fan},
title = { { DDecomp: Structured Uncertainty Decomposition Using Masked Autoencoders for Anomaly Detection in Diffusion MRI } },
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 DDecomp, an unsupervised anomaly detection framework for diffusion MRI that is trained only on healthy data and detects brain lesions through uncertainty decomposition. The main idea is to separate aleatoric uncertainty, which comes from acquisition noise and cross-protocol variability, from epistemic uncertainty, which is more indicative of out-of-distribution pathological tissue. The method then uses the epistemic component as a more pathology-specific signal for lesion localisation. The method also marginalises dropout and mask sampling to obtain a more stable uncertainty estimate, and is evaluated across multiple lesion types and acquisition protocols.
- 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.
This paper addresses an important and challenging problem: unsupervised anomaly detection in diffusion MRI under substantial acquisition noise and cross-protocol variability. This method is designed specifically for this problem rather than relying on reconstruction error or a generated anomaly score. The method formulation is technically sound and clearly structured, combining a healthy-data-trained masked autoencoder with heteroscedastic likelihood modelling and inference-time marginalisation over mask sampling and dropout to produce a more stable and pathology-specific uncertainty estimate.
In addition, the evaluation across multiple lesion types (glioma, WMH, and ICH) and protocols shows a superior performance of the proposed method, which strengthens the practical relevance of the work.
- 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.
One concern relates to the choice of using dMRI-derived scalar map mean diffusivity (MD) as model input. The authors already mentioned that reliable lesion localisation in scalar diffusion maps remains challenging in the introduction. MD can be sensitive to various artefacts. This is an important limitation, because these derived measures are influenced not only by pathology but also by acquisition quality, preprocessing errors, and protocol variability. In practice, severe artefacts or unstable diffusion estimation may also produce abnormal intensity patterns that resemble pathology, meaning the anomaly signal may not be lesion-specific. As a result, the proposed framework may detect generic deviations from the healthy training distribution, including artefacts, noise contamination, or reconstruction failures, rather than truly pathological tissue alone. Therefore, an elevated epistemic uncertainty map may reflect either a clinically meaningful lesion or simply poor image quality. The paper would be stronger if it included a more explicit discussion or analysis of how they perform on the artefact dataset.
Another weakness is the lack of clarity regarding the evaluation datasets and reproducibility. The manuscript reports experiments on glioma, intracerebral haemorrhage (ICH), and white matter hyperintensity (WMH) cohorts, but does not clearly identify the source of these datasets, provide corresponding references, or state whether they are publicly available or in-house collections. The authors also do not provide information regarding the demographics of these datasets. This makes it difficult to assess the diversity, accessibility, and potential biases of the evaluation data, and limits the reader’s ability to judge how broadly the reported results may generalise. In addition, the paper does not appear to provide any statement regarding code availability. For a method that involves several implementation-specific components, including masked autoencoding, heteroscedastic uncertainty modelling, and inference-time marginalisation, the absence of released code further reduces reproducibility. Overall, the limited transparency around both the datasets and the implementation weakens the practical reproducibility of the work.
- 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 provide sufficient information for 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 paper has a clear technical idea, and the overall method is reasonably well motivated for unsupervised anomaly detection in diffusion MRI. In particular, the decomposition of predictive uncertainty into aleatoric and epistemic components is a meaningful contribution, as it attempts to distinguish acquisition- and protocol-related variability from pathology-related model mismatch. The methodological formulation also appears technically sound, with a healthy-data-trained masked autoencoder, heteroscedastic modelling, and inference-time marginalisation over masking and dropout to stabilise uncertainty estimation. In addition, evaluating a single model across multiple lesion types is a positive aspect and suggests potentially useful generalisability.
However, I do not find the paper strong enough for a higher score because several important concerns remain. First, the reliance on diffusion-derived scalar maps (MD as input) raises questions about lesion specificity, since severe artefacts, preprocessing failures, or protocol-related abnormalities may also appear as anomalies. This makes it unclear whether the resulting epistemic maps truly isolate pathology or more broadly reflect deviations from the healthy training distribution. Second, the paper lacks transparency regarding the experimental data: the sources of the glioma, ICH, and WMH datasets are not clearly identified, references are missing, and it is not stated whether the cohorts are public or in-house. Third, there is no clear statement regarding code release, which limits reproducibility for a method with several nontrivial implementation details.
Overall, I think the paper is technically interesting and likely above the rejection boundary, but the current presentation leaves some uncertainty about robustness, specificity, and reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The authors did not reply to my concerns about the usage of MD as input; however, I believe it is still an important choice to justify for dMRI.
Review #2
- Please describe the contribution of the paper
Anomaly detection of brain lesions is critical for supporting neurological disease diagnosis and surgical planning. This study proposes DDecomp, which is based on masked autoencoders trained on healthy data and decomposes predictive uncertainty into the epistemic component for unsupervised anomaly detection. The method was validated on different datasets for different lesion types.
- 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 decomposition of the epistemic component for unsupervised anomaly detection is interesting
- Different lesion types have been included for comprehensive evaluation
- 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 methods considered for comparison were too limited
- The selected evaluation metrics were not sufficient
- 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.
(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 method is not necessarily restricted to diffusion MRI. It can be directly applied to other MRI modalities. Please discuss.
-
The claim that the epistemic component is more suitable for anomaly detection should be supported with experimental evidence.
-
The major concern about this work is the lack of comparison with state-of-the-art methods. There are numerous unsupervised anomaly detection methods that can be applied to the tasks in this paper. For example, https://doi.org/10.1016/j.media.2019.01.010, https://proceedings.mlr.press/v172/kascenas22a.html, https://doi.org/10.1007/978-3-032-04965-0_43, https://doi.org/10.1007/978-3-031-96625-5_1, and many others.
-
To be clinically useful, the anomaly scores need to be converted to binary lesion masks. However, the proposed method does not address this problem. There are no Dice coefficients or distance measures available to evaluate the accuracy of lesion boundaries. The clinical relevance remains to be justified.
-
Since MC sampling is used, the computation efficiency should also be evaluated.
-
- Reviewer confidence
Very confident (4)
- [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 two major factors that led to my original recommendation are not addressed: 1) It is not clear how F3 is computed and what it represents. The selection of \beta=3 is not justified. Also, threshold selection is challenging, yet it is unknown how it was performed for each method. Thus, it is still not convincing that this method can lead to good segmentation accuracy. 2) The two methods selected for comparison were insufficient. They are not the common ones used for comparison by existing literature.
Therefore, the advantage of the proposed method has not been adequately demonstrated by the experiments, and I stick to my original recommendation.
Review #3
- Please describe the contribution of the paper
The paper describes a method for mapping anomalous / out-of-distribution regions in images after learning from a population of images from ostensibly healthy subjects. The paper focusses specifically on diffusion MRI and uses the HCP data set for training. The authors aim to separate aleatoric uncertainty (from noise etc) from epistemic uncertainty (inherent to the data) and map regions of high epistemic uncertainty as indicative of signals that do not arise in the training data and thus likely to represent pathological tissue.
- 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 idea has good novelty (at least I haven’t seen normative modelling and abnormality detection done this way)
-
I particularly like the idea of inferring the posterior on the mask by treating it as a nuisance variable.
-
- 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.
-
I don’t fully understand why high epistemic uncertainty necessarily corresponds to abnormality. There may be very common signal combinations for which epistemic uncertainty is high (e.g. where many signal combinations map to the same diffusion parameters).
-
Not necessarily a weakness but a question: the HCP data set is predominantly “healthy” but pathologies certainly exist and are visible in the images (the website provides lists and locations). Are these excluded from training or is the training robust to modest inclusion of “abnormal” signals? If the trained model were then applied to the training data, does it pick up such abnormalities?
-
The Discussion is thin. No mention or acknowledgement of limitations / further work etc, which should be corrected.
-
- 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.
(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?
I’m not 100% on this as there are certainly queries, but I like the idea and the work seems to have high novelty.
- 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.
I still like this work. It does have some limitations but some nice ideas in there too so I think worthy of acceptance. I’m leaving my score as it is.
Author Feedback
Uncertainty Separation (R1–3, AC): We apologize for confusion over how to separate lesions from image artifacts. Our method first estimates total uncertainty, then decomposes it into epistemic and aleatoric components via variance analysis of MC sampling and heteroscedasticity modeling. Epistemic uncertainty (measured as variance of predictive means across MC-dropout and masked samples) captures OOD from the healthy training distribution, e.g., pathological lesions with rare anatomical patterns. Image artifacts (e.g., protocol shifts, noise, and signal combinations) also shift test data from the training data characteristics, but these commonly exist across visible and masked tokens in MAE during inference, which will be captured by our heteroscedastic likelihood computation. Thus, the two uncertainties can be separated for epistemic lesion identification. Our comparison with MAE-UAD and DDEvENet shows that separating these uncertainties largely improves the identification. But, we acknowledge that our method relies on high-quality processed data and QC to exclude failures. Otherwise, severe imaging artifacts (e.g., reconstruction failures) may be recognized as epistemic OOD.
Segmentation Accuracy & UAD Comparison (R2, AC): (1) We will clarify that Fβ (Tab. 1) quantifies segmentation accuracy by measuring the similarity between binary lesion masks (by thresholding the anomaly scores) and manual segmentation. (2) Due to space limit, our current work compared two most relevant approaches: MAE-UAD (a reconstruction-error method from the same MAE family as ours) and DDEvENet (an uncertainty estimation method designed for dMRI on the same task and data as ours). We agree that including more UAD methods will strengthen our evaluation, and will expand to discuss these work and include comparisons in future journal extension.
Abnormalities in Training (R3, AC): We confirm that no cases were excluded due to pathologies during training, and acknowledge that incidental abnormalities may be present when learning the healthy population normative model. As such cases are rare and spatially sparse relative to normal whole‑brain tissue, they will cause limited contamination rather than deviating the learned normative model. Moreover, our random masking design further prevents such contamination because abnormal regions form a small fraction of visible token targets. The heteroscedastic likelihood will also assign higher variance to such isolated high-residual observations, reducing their effects on the learned mean. Identifying such cases in the training set would require further investigation into abnormal data characteristics, which is beyond the scope of this work. But we believe these abnormalities would be detected as they lie outside the learned normative distribution.
Data Transparency and Code (R1, AC): Due to the anonymous policy, we did not include source of our in‑house clinical data. Upon acceptance, we will add data details, including demographics and clinical characteristics. Further, we confirm that the code and trained weights will be released.
Beyond dMRI (R2, AC): AC and R2 are correct that our method can be generally applied to other modalities. We will clarify that it is a general uncertainty-decomposition framework, not task-specific. This work focuses on dMRI as it’s a natural application for separating aleatoric variability from epistemic OOD response due to its inherent variation from acquisitions, scanner differences, and microstructural pathology.
Computation Efficiency (R2): Our method requires K (# of random masks, 100) x T (# of MC-dropout samples, 10) forward passes for inference, which takes ~2 min per subject.
Expanded Discussion (R3): As suggested, we will expand Discussion to include limitations regarding severe artifact effects, reliance on high-quality processed data, potential abnormalities in the training data, and future work on testing beyond dMRI as well as comparisons with more UAD methods.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
This manuscript proposes a masked autoencoder based framework for anomaly detection in diffusion MRI that decomposes predictive uncertainty into aleatoric and epistemic components. Reviewers thought that this was an interesting and innovative approach to a relevant and challenging problem, and yielded promising qualitative results. However, they also raised a number of concerns that are too serious to directly recommend acceptance. It would be helpful if authors could comment on them in their rebuttal, keeping in mind that no new results are permitted at this point: 1.All reviewers questioned how helpful and reliable the separation between aleatoric and epistemic uncertainty really is for separating lesions from other abnormalities such as image artifacts. They thought that this was not demonstrated clearly enough. 2.The accuracy of the proposed approach as a lesion segmentation tool has not been quantified, and reviewers missed a comparison to multiple plausible baselines for anomaly detection. 3.It is unclear how abnormalities in the HCP training data have been handled. 4.The sources and characteristics of the clinical data used in this work are not sufficiently transparent. 5.Reviewers expressed doubt that, without making source code available, it will be possible to reproduce this work. 6.It is not clear why the authors focus on diffusion MRI, since the method should be more widely applicable.
- 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.
Some limitations remain after rebuttal and should be clearly stated in the final version of the paper, in particular, assumptions regarding image quality that underlie the proposed separation between aleatoric and epistemic uncertainty, and assumptions regarding effects of training data contamination being negligible. Concerns regarding reproducibility have been addressed by promising the release of code and model weights along with the paper. Authors are encouraged to include additional information about selection of anomaly score threshold and stability of results w.r.t. this choice, which is a major remaining concern. Despite this, I agree with the positive reviewer majority vote after rebuttal.
Meta-review #2
- 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 uncertainty decomposition idea is genuinely novel and the dMRI application is well-motivated. However, the UAD field offers a substantial number of directly applicable methods and situating the contribution in this context is not optional. Adding baseline comparisons requires only additional rows in a table and deferring this to a journal extension is not an appropriate rebuttal response. The claimed improvements cannot be properly assessed without this context. Additionally, no dataset information is provided beyond lesion type, including sources, demographics, or acquisition protocols. This is feasible even under anonymity and its absence makes it impossible to assess generalizability, site bias, or protocol dependence for a method that explicitly claims robustness across these factors.
Meta-review #3
- 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.
I am choosing to accept this paper, but I do agree with one of the points raised by R2, that the experimental baselines are disappointing. Still, it is innovative and interesting, and with 2 vs 1 it becomes editorial choice.
