List of Papers Browse by Subject Areas Author List
Abstract
Clinical magnetic resonance imaging (MRI) scans often exhibit substantial variability in slice thickness owing to heterogeneous imaging protocols and scan-time constraints. However, most deep learning models for brain MRI are developed under fixed-resolution assumptions, limiting their generalization across resolution shifts. To fill the gap, we propose a resolution-conditioned representation learning framework that can effectively handle scans with varying slice thicknesses. Our framework is pretrained via self-supervised reconstruction on 21,156 brain MRI scans from 14 public cohorts, and the learned representations are further refined through multimodal alignment with clinical metadata, including acquisition parameters and demographic characteristics. This alignment enables the model to accommodate diverse acquisition settings while preserving fine-grained anatomical information. The learned representations consistently support multiple clinical applications, such as brain tissue segmentation and diagnosis of autism spectrum disorder and major depressive disorder.
Extensive experiments on public benchmarks and in-house clinical cohorts demonstrate that our framework achieves competitive segmentation performance against state-of-the-art models and exhibits superior generalization to external datasets. Notably, our method outperforms task-specific diagnosis models on real-world thick-slice MRI scans, highlighting its robustness under clinically realistic imaging conditions. These findings demonstrate the potential of resolution-conditioned representation learning to enable robust and versatile brain MRI analysis across heterogeneous slice thicknesses.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1434_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
ABIDE dataset: https://fcon_1000.projects.nitrc.org/indi/abide/
ADNI dataset: https://adni.loni.usc.edu/
OASIS-3 dataset: https://sites.wustl.edu/oasisbrains/
UK Biobank: https://www.ukbiobank.ac.uk/
ABCD Study: https://abcdstudy.org/
HCP-YA: https://www.humanconnectome.org/study/hcp-young-adult
HCP-D: https://www.humanconnectome.org/study/hcp-lifespan-development
Additional public datasets are described in the manuscript.
BibTex
@InProceedings{DonKe_ResolutionConditioned_MICCAI2026,
author = { Dong, Ke AND Hu, Xihong AND Huang, Shijie AND Xu, Yanqing AND Lian, Zifeng AND Li, Xiaoye AND Yang, Liqin AND Fan, Yimin AND Shen, Dinggang AND Sun, Kaicong},
title = { { Resolution-Conditioned Representation Learning for Unified Brain MRI Modeling } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a novel method for representation learning for structural brain MRI. The method combines many modules that are not new, but they are logical choices to solve the given problem. The method was tested on downstream tasks of segmentation, disease classification, and image retrieval.
- 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 adopted modules are logical choices and the method was tested on many downstream tasks.
- 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.
- Weak comparison. The introduction mentions OmniRad as a related method, which was not compared. There are recent methods like <A generalizable foundation model for analysis of human brain MRI, Nature Neuroscience 2026> and <A foundation model for enhancing magnetic resonance images and downstream segmentation, registration and diagnostic tasks, Nature BME 2025>. These should be mentioned in the Introduction and compared against in the Results.
- Weak ablation. The method is largely applicable to any image backbone. The results are shown only for one CNN-based backbone. Results across different backbones are needed.
- During inference, the low-resolution data are resampled to 1x1x1 and then applied. There should be a way to directly incorporate low-resolution input.
- It is unclear what kind of clinical variables are used. They should be explicitly stated. It is unclear how contrastive learning is performed. Is it across the same vs. different patients? Is it across the same vs. different imaging setup? Why do you finetune only the final residual block for the contrastive learning?
- The testing should include real low-resolution datasets for segmentation. Not simulated low-resolution cases.
- Table 4 lacks p-values and recent baselines. You cannot just compare with one baseline of ViT.
- Table 5 lacks p-values.
- Procedures to perform retrieval are unclear.
- 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.
(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?
I am very concerned with weak comparison (#1) and weak ablation across backbones (#2). I am also concerned with limited testing (#5) and insufficient statistical comparisons (#6, #7). This led to my rating.
- 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.
Some issues are clarified. Still, weak comparative results (across recent methods, backbones) persist.
Review #2
- Please describe the contribution of the paper
The paper proposes a resolution-conditioned representation learning framework for brain MRI that explicitly models slice-thickness variation, using large-scale reconstruction pretraining plus cross-modal alignment with clinical metadata.
It is pretrained on 21K brain MRI scans from 14 public cohorts and evaluated on multiple downstream tasks, including tissue segmentation, psychiatric disorder classification, and image-to-text retrieval, across public benchmarks and in-house clinical datasets.
The method achieves competitive performance under mixed-resolution training, improves robustness on unseen-center data, and outperforms task-specific models on real-world thick-slice MRI.
- 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 paper represents an interesting take on using self-supervision and weakly supervised learning to train stronger encoders for brain MRI. In this, the paper stands out in its simplicity, while seemingly obtaining strong performance, in particular on clinical grade scans.
-
The problem of training encoders for anisotropic data is highly relevant for applying models to real clinical data and in this regard, the comparison to SynthSeg (the current sota model on clinical scans) is appreciated.
-
The evaluation is strong, with relevant baselines.
-
The ablation study, ablates the most central design choices, and thus provides important insights into why and how the method works, however a few central ablations are missing (see weaknesses).
-
The paper is generally well written.
-
- 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 paper lacks important information on the resolution robust representation learning stage: was the data resampled both before and after slice-corruption, with the first used as ground truth in the reconstruction task?
-
It is unclear if the method works on pretraining datasets containing anisotropic data. In the proposed setup, an anisotropic base image, resampled to isotropic resolution, seems to be a poor ground truth for reconstruction. This remains a clear limitation, since the application of the method to clinical archives or recent datasets such as OpenMind (120K scans) and FOMO300K (300K scans), is limited.
-
Since the method applies reconstruction in both stage 1 and stage 2 (combined with a contrastive loss for metadata alignment), the exact effect of each reconstruction step seems important to examine. Could the method skip stage 1 and obtain similar performance? What happens if we dont do reconstruction during stage 2?
-
The paper is missing two key sections: related works and limitations. In particular, the paper is poorly situated in the literature missing key comparisons to previous work, for instance [1] which also seeks to learn encoders which work well on anisotropic data and [2] which does contrastive alignment of brain mri scans and metadata applied to disease diagnosis and image-to-text retrieval.
-
The paper does not promise to release code or weights, which significantly hinders reproduction of the study results.
-
A key obstacle when doing contrastive learning in 3D is the small batch sizes possible in 3D. The paper lacks details on how this was overcome and the method’s stability over different batch sizes.
[1] https://arxiv.org/abs/2410.08397 [2] https://arxiv.org/abs/2501.14051
-
- 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
Please provide clarifications on the points raised in the weaknesses section.
- 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 tackles a highly relevant problem – learning MRI encoders that remain effective on anisotropic, clinical-grade data – with sound evaluation and generally informative ablations. The method is also appealing in its simplicity and appears to deliver practically meaningful gains, which gives it real potential impact. That said, the paper is held back by missing methodological detail, insufficient analysis of the two reconstruction stages and poor positioning with respect to related work. I am willing to raise my score if my concerns are adequately med during the rebuttal phase.
- Reviewer confidence
Very confident (4)
- [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 rebuttal has adequately responded to my questions and i thereby recommend acceptance.
Review #3
- Please describe the contribution of the paper
This work presents a resolution-conditioned representation learning framework designed to explicitly address the variations in slice thickness in clinical brain MRI scans. By pretraining a 3D ResNet-50 encoder on a large-scale collection of 21,156 brain MRI scans using a reconstruction-based self-supervised objective, and integrating structural clinical metadata (e.g., age, gender, imaging parameters) via contrastive alignment, the model learns resolution-aware and anatomy-preserving features. This unified framework is shown to be effective across diverse downstream tasks, including brain tissue segmentation, disease classification (ASD and MDD), and image-to-text retrieval.
- 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 paper tackles a practical and clinically relevant problem that is often under-addressed in brain MRI modeling, namely robustness to heterogeneous slice thickness. 2.The framework is coherent and has a reasonably broad scope, covering multiple downstream tasks rather than optimizing for a single benchmark. 3.The use of acquisition metadata for multimodal alignment is a sensible design choice and clinically realistic, since such information is routinely available. 4.The experimental section looks reasonably broad, with internal and external segmentation evaluation, disease classification, and ablations.
- 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 methodological novelty is not enough. The paper puts together several reasonable components, but the conceptual advance over prior representation-learning pipelines is limited. 2.The segmentation gains over strong baselines are relatively small, so the practical significance of the improvement should be discussed more carefully. 3.The strongest robustness claims rest mainly on simulated thick-slice experiments rather than truly acquired thick-slice segmentation data. For a paper centered on clinical robustness, that is an important limitation; the author should explore this limitation. 4.The image-to-text retrieval experiment does not add much to its current result, as the reported retrieval performance is modest and not benchmarked against clear baselines. 5.The proposed pipeline assumes all scans are skull-stripped and rigidly registered to the MNI template. In practical, heterogeneous clinical settings, such preprocessing might fail or be suboptimal on thick-slice data, which could degrade the downstream performance. This influence should be discussed.
- 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?
Overall, I think this is a solid paper on an important practical problem, and the proposed framework is technically sound and fairly well evaluated. My concerns are mainly about the level of novelty, the reliance on simulated rather than real thick-slice evidence for the central robustness claim, and the limited disease-classification comparisons.
- 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 author has addressed most of my concerns. So I changed my decision.
Author Feedback
We thank the meta reviewer and reviewers for your positive and constructive comments. We address the main concerns below. R1/2: Comparison with related foundation-model studies. We acknowledge that comparison with related foundation models would strengthen the positioning of the work. Since one of our main contributions is model robustness under heterogeneous slice thickness, in preliminary evaluation, we actually have compared with the foundation model BrainIAC, but we found its performance under heterogeneous slice thickness is worse (Dice: 0.83 at 6 mm) than task-specific models (>0.89). Another foundation model VoxelPrompt, which was designed for anisotropic volumes, can only provide output with the same dimension as the input. However, our method is used for thick-to-thin cases, so it cannot be directly compared with our method. Therefore, in previous version, we only compared with task-specific benchmarks due to limited page. In future work, we will compare with these foundation models. R1/2/3: Real vs. simulated thick-slice robustness. In fact, our paper already includes evaluation on real clinical thick-slice MRI for downstream ASD/MDD diagnosis tasks. Regarding segmentation task, since controlled slice-thickness simulation provides a cheap yet effective option as shown in prior studies [1,2] while real-world paired data is rare and needs manual annotations, we chose to conduct experiments on simulated data and our model achieves 0.88 in Dice on 6-mm cases, showing strong confidence for its application to real thick-slice segmentation. In future work, we will collect real-world data pairs and perform manual labeling for clinical evaluation. [1] Billot, B. et al., SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining, Med Image Anal, 2023. [2] Ayaz, A. et al., Brain MR image simulation for deep learning based medical image segmentation, JBHI, 2023. R1/2: Clarification of the staged multimodal training strategy. The staged design balances semantic alignment and anatomical fidelity in 3D multimodal learning. We found simultaneously optimizing reconstruction and contrastive objectives during the pretraining cause too much memory on 3D medical data, limiting the batch size maximum to 8 and resulting in less effective contrastive training. We additionally observed that removing reconstruction regularization during Stage 2 can increase the contrastive batch size, but degrade downstream segmentation quality, suggesting that contrastive learning may favor semantic shortcuts at the expense of fine anatomical structure. Therefore, we used Stage 1 to establish stable anatomy-preserving representations, and Stage 2 to perform semantic adaptation by fine-tuning only the final residual block, which supports larger batch size up to 16.R1/2/3: Additional clarifications. Regarding different backbones, actually we have explored transformer-based 3D SwinUNETR (feature size=48). Due to its high memory footprint in 3D multimodal training, the feasible batch size was limited to 8.Under this batch size, contrastive learning for Stage 2 was less effective, with retrieval Top-1 performance dropping to 18%. We therefore adopted ResNet-50 as a more stable evaluation setting. We will evaluate more backbones in the journal work. Regarding statistical significance, for classification in Table 4, performance gain in AUC by our method is consistently greater than 6% across three datasets. Hence, we ignored to show the p-values in our previous version (p < 10-32). For ablation in Table 5, each step shows Dice improvement of at least 1%, which is also statistically significant (p < 10-18). Regarding isotropic resampling, we aim to establish a common feature space across heterogeneous protocols for more effective representation learning, which is actually standard for downstream tasks such as AD diagnosis. In revised version, we will provide more details of method descriptions and contrastive learning setup.
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.
The reviewers identify several critical areas needing attention in the rebuttal, focusing primarily on the paper’s robustness claims and comparative analysis. The most significant concern is the weak comparison against recent foundation models and an over-reliance on simulated, rather than real, thick-slice data, which undermines the claim of real-world efficacy. Hard to see the source of performance gains when there is no clear ablation across different backbones. The authors must also provide essential methodological details and the contrastive learning setup and why there are two reconstruction stages.
- 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.
Following the rebuttal, the reviewers agree that the authors addressed the initial concerns regarding baseline comparisons and ablation studies. Given the framework’s practical clinical relevance in handling heterogeneous slice thicknesses, I recommend the paper for acceptance.
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.
After carefully reviewing the authors’ rebuttal, I believe that the authors have effectively addressed the concerns raised by all reviewers. The responses are clear and sufficiently detailed, and the revisions have improved the manuscript accordingly. Therefore, I recommend acceptance. Ins
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.
After careful review of the paper and the rebuttal, several major problems prevent me to recommend acceptance.
Most important is an issue of score manipulation. SynthSeg is one of the SOTA in brain segmentation: the original paper and many subsequent studies have shown performance around 80/90% Dice scores on 1mm data. This is very far from the 66% Dice scores shown here on the same/similar datasets as previous studies. Therefore, these scores are not credible at all, which questions all the results presented in this submission.
Moreover, based on my clinical experience, relying on clinical metadata at inference time is not a good strategy, because this data is very often corrupted/erroneous or simply missing.
These issues combined with other problems raised by reviewers about incremental novelty, insufficient description of the methods and experimental setup, missing baselines in SOTA brain foundation models, are fatal to recommend acceptance. While clarifications about methods and setup were given during rebuttal, novelty and missing baselines remain a primary concern.
