List of Papers Browse by Subject Areas Author List
Abstract
Structural MRI (sMRI) has revealed morphological correlates associated with psychiatric disorders, motivating neuroimaging-based diagnostic support tools. However, multi-class classification from sMRI remains chal-lenging in small-cohort psychiatric settings, where site-level datasets often contain only a few hundred subjects and exhibit class imbalance, and where subtle inter-class differences further complicate discrimination. We propose a class-aware representation learning framework (CARformer) that adapts domain-pretrained encoders for 3D sMRI multi-class psychiatric classifica-tion. The framework includes (1) a class-aware Bag-of-Queries (CA-BoQ) module that learns class-aligned query representations to aggregate class-specific dis-criminative evidence under inter-class differences, and (2) a Class-Aware Contrastive (CAC) learning strategy that strengthens class separation under imbalance. We evaluated the framework across two sites from a public brain MRI dataset to assess robustness under differences in diagnostic composition, class imbalance, and imaging acquisition proto-cols. Our results demonstrate consistent improvements in macro AUC and AUPRC over existing standard 3D Convolutional Neural Network (CNN) backbones and domain-pretrained fine-tuning frameworks across both sites, supporting the practical utility of class-aware adaptation of domain pre-training in small-cohort, imbalanced psychiatric imaging tasks. Source code is available at: https://github.com/xyfu33/CARformer.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3161_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xyfu33/CARformer
Link to the Dataset(s)
SRPBS Multi-disorder MRI Dataset: https://bicr-resource.atr.jp/accounts/create_srpbsopen/
BibTex
@InProceedings{FuXin_CARformer_MICCAI2026,
author = { Fu, Xingyue AND Deng, Ruoxi AND Ahn, Euijoon AND Korgaonkar, Mayuresh S. AND Kim, Jinman},
title = { { CARformer: Class-Aware Representation Learning for Small-Cohort, Imbalanced Psychiatric Disorder Classification in Brain 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
The authors propose a novel framework, termed CARformer, to address the challenges of small-sample, class-imbalanced, and multi-class classification in psychiatric sMRI analysis. Specifically, the method introduces two novel class-aware components: the CA-BoQ module for class-specific feature aggregation and the CAC strategy for confusion-aware contrastive learning. Through evaluation on two independent clinical sites, the paper shows that the proposed approach outperforms several baseline methods and substantially improves the recognition of minority classes.
- 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.
- Well-motivated problem setting: The study targets a highly relevant and practically important clinical challenge, namely small-sample, imbalanced, and multi-disease classification in psychiatric neuroimaging.
- Methodological novelty: The “class-aware” design is conceptually interesting and appears well tailored to the problem setting.
- Rigorous experimental evaluation: The method is assessed on two independent and heterogeneous clinical sites, which strengthens the robustness of the findings and suggests good generalization potential.
- Promising empirical performance: The reported improvements are clear, especially for minority classes, with particularly notable gains on metrics such as AUPRC.
- 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.
- Insufficient methodological details: Critical implementation and design details are missing, which substantially weakens the rigor and reproducibility of the work. For example: (a) it is unclear how the logits from the CA-BoQ and the GAP are combined (e. g. , direct summation or weighted fusion), and why this design was chosen; (b) the initialization, update frequency, and update strategy of the confusion matrix W in the CAC module are not defined, leaving an important part of the method underspecified; (c) the manuscript does not discuss how the stability of the class-aligned queries is ensured during training.
- Lack of control for potential confounders: The paper does not provide sufficient evidence that the diagnostic groups are comparable at baseline, nor does it clearly state whether and how key covariates such as age and sex were controlled for in the training or analysis pipeline.
- 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 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 presents a meaningful methodological idea and promising empirical results, particularly for minority-class recognition. However, the current manuscript requires substantial revision before it can meet the standard for publication, mainly because the description of the method lacks essential details, which raises concerns regarding rigor 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’ response is comprehensive and substantive. They have directly provided the key methodological details that were previously missing, making the overall rationale, implementation, and reproducibility of the CARformer framework much clearer. Regarding the control of potential confounding factors, the authors also offered a reasonable explanation that is consistent with common practices in clinical neuroimaging research. Based on these revisions and responses, I believe that the manuscript has now reached an acceptable level for publication.
Review #2
- Please describe the contribution of the paper
The authors build on a domain-pretrained 3D Uniformer encoder and propose CARformer, which combines a Class-Aware Bag-of-Queries module for aggregating class-specific discriminative representations for classification and a Class-Aware Contrastive learning strategy for transfer learning. Overall, the paper presents a relatively complete class-aware representation learning framework and explores its potential for more objective diagnostic support in psychiatric disorders.
- 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 motivation is well grounded. The paper tackles the problem under realistic clinical constraints, including small sample size, class imbalance, and limited single-site data, which makes the problem itself practically meaningful. 2.The dual-branch design is reasonable. The combination of global average pooling and CA-BoQ helps balance training stability and discriminability, which is a sensible design choice for this task.
- 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 evidence for generalizability is insufficient. The experiments are conducted only on two sites from a single dataset, both from Japanese cohorts. There is no cross-dataset or cross-ethnicity validation, which limits the evidence for broader generalization. In addition, the set of baseline methods is relatively limited and somewhat outdated. 2.Important implementation details are missing. In particular, the paper does not clearly explain how the confusion matrix W is updated or how the corresponding weights are computed, which affects reproducibility and makes the CAC component harder to assess. 3.The interpretability claim is not sufficiently supported. The attention maps in Figure 2 do not appear to indicate stable and reproducible disease-related biomarkers across sites. This weakens the authors’ claims regarding spatial interpretability and biological consistency.
- 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.
(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 paper does not yet demonstrate sufficient practical utility, and the experimental evidence is not strong enough to support its broader claims. In particular, the evaluation is limited to two sites from a single dataset, without cross-dataset or cross-ethnicity validation, and the set of baselines is relatively limited. As a result, the current study is insufficient to establish the method’s robustness and real-world applicability.
- 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.
After reading the rebuttal, I am satisfied with the authors’ clarification. My initial concern was partly due to my limited familiarity with this specific dataset and task setting. The authors’ response helped clarify the data usage and the motivation behind their experimental design. I now find their explanation reasonable and believe that the main concern I raised has been adequately addressed. Therefore, I am willing to revise my assessment in a more positive direction.
Review #3
- Please describe the contribution of the paper
The authors propose a class-aware representation learning framework for multi-class psychiatric disorder classification from brain sMRI under small-cohort, imbalanced data distributions.
- 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 authors propose a novel CARformer framework that integrates two complementary class-aware components to address multi-disorder discrimination under small-cohort, imbalanced psychiatric sMRI settings, which improves inter-class separability through class aware supervised contrastive supervision with confusion-informed negative re-weighting.
- 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 introduction mentions co-morbidities and suggests that the methodology would be beneficial in determining overlap. However, the rest of the paper follows a traditional 3 group differential diagnosis path, all-be-it with the capability of handling different group sizes.
There is no real explanation as to why the accuracy for HC or MDD is significantly different between hospitals. Given the proposed method is only better in 50% of the measures in the KUT dataset it suggests that CARFormer is not so much better than the other methods tested.
- 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.
(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 methodology does not seem to provide a significant improvement in accuracy compared to previous methods and there is no clear pathway to clinical application. Multimorbidity of psychiatric and neurodevelopmental disorders is an important area for research and this method may provide a useful tool in the future, but at present that has not been proven.
- 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 have adequately responded to my comments.
Author Feedback
We thank the Reviewers(Rs) and AC for their constructive feedback and for recognizing CARformer’s clinical motivation, class-aware methodological novelty, minority-class performance gains, and dual‑site evaluation in real‑world data conditions. We address the key concerns below and will add these clarifications and code-release information in the revised manuscript. Methodology Clarification: (R1 Q7.1a) CA-BoQ and GAP logits are fused by weighted additive logit blending, z=z_GAP+ α*z_BoQ, with fixed α=0.3.GAP provides a stable global classifier from the pretrained encoder, while CA-BoQ uses K learnable class-aligned queries and a diagonal per-query head to produce class-specific logits. (R1 Q7.1b; R2 Q7.2) We clarify that the CAC confusion matrix W is updated only from validation predictions. Before the first validation pass, CAC uses uniform negative-pair weighting with W_ij=1.After each validation epoch, W is computed from the row-normalized validation confusion matrix, clipped to [1,2], with diagonal entries reset to 1; subsequent epochs use exponential smoothing with decay = 0.5.In CAC, W_ij reweights only negative pairs between classes, while positive pairs remain as weight = 1.(R1 Q7.1c) Class-query stability is supported by both architectural and regularization designs: (i) class-aligned query slots with LayerNorm before/after cross-attention and a shared diagonal class-query head; (ii) a query-diversity penalty (λ_div = 10^-3) to reduce query collapse. CAC operates GAP-pooled features rather than query embeddings. Confounder handling (R1 Q7.2): Table 1 reports diagnostic-group demographics and scanner protocols for each SRPBS site (Tanaka et al., 2021). Although the cohort is not fully matched across diagnostic groups at baseline, potential confounding was mitigated through site-specific evaluations and subject-level stratified train/validation/test splits, preserving diagnostic distributions and balancing age and sex across subsets. Generalizability & baselines (R2 Q7.1) a) As a primarily methodological contribution, our experimental design was necessarily scoped by the availability of data. We note that the dataset we used is currently the only publicly available resource for the research community to study this clinically important problem. We will better justify this scope and explicitly discuss cross-dataset and cross-ethnicity limitations. b) Our baselines include 3D CNNs used for psychiatric sMRI classification (Zhang et al., 2023), BrainMVP (Rui et al., CVPR 2025) – a brain MRI pretraining model evaluated on brain MRI classification tasks and benchmarked against training-from-scratch and general/medical-domain SSL methods; and PCRLv2 (Zhou et al., TPAMI 2023), a strong medical-domain SSL framework. These represent task-relevant 3D sMRI classifiers and pretrained medical/brain MRI fine-tuning comparators, rather than only generic CNNs. Attention map (R2 Q7.3) UTO and KUT involve different disease contrasts, HC/MDD/BD versus HC/MDD/SZ, Fig. 2 visualizes CA-BoQ’s class-specific attention within each differential-diagnosis task, rather than testing invariant disease biomarkers across sites. Because competing diagnoses differ, discriminative evidence for the same class, such as MDD, may also differ across sites. Site performance differences (R3 Q7.2) We like to respectfully suggest that HC/MDD metrics should not be interpreted as the same binary task repeated across hospitals; MDD is contrasted with BD in UTO but with SZ in KUT, and sites differ in class distributions and acquisition protocols. CARformer is designed to improve class separation, in highly imbalanced KUT setting, it achieved best macro AUC/AUPRC and substantially improving minority-class MDD AUPRC from 0.196 with ResNet18 (the second-best macro baseline) to 0.700.Comorbidity/overlap clarification (R3 Q7.1) We clarify that comorbidity/symptom overlap is used as clinical motivation for the difficulty of 3-class psychiatric differential diagnosis.
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 decision is grounded in the work’s clinically well-motivated research target on small-cohort imbalanced psychiatric disorder classification, the methodological novelty of the proposed CARformer framework, and valid dual-site experiments showing clear improvements for minority classes, alongside mixed but marginally competitive overall reviewer scores. To facilitate fair further evaluation, the authors must primarily address these key points: complete missing implementation details of CA-BoQ, GAP fusion and the confusion matrix update strategy, clarify unregulated clinical confounders like age and sex, respond to limited cross-dataset generalization and outdated baseline settings, justify unstable attention map interpretability and explain inconsistent model performance across experimental sites.
- 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.
All three reviewers updated their final recommendations to Accept after reading the authors’ comprehensive rebuttal, which clarified key methodological details and improved the rationale and reproducibility of the CARformer framework. They found the authors’ explanation of confounding factor control aligned with common practices in clinical neuroimaging research and acknowledged that their initial concerns were partly due to a lack of familiarity with the dataset and task setting. Overall, they believe the authors’ responses have adequately addressed their comments.
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.
Authors have adequately addressed all reviewers’ concerns. There is a consensus toward acceptance, which I agree with.
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.
Based on the three post-rebuttal reviews, all reviewers are satisfied with the authors’ responses. Reviewer #1 finds the response comprehensive and substantive, noting that the manuscript has reached an acceptable level for publication. Reviewer #2 acknowledges that their initial concern has been adequately addressed and revises their assessment positively. Reviewer #3 confirms that the authors have adequately responded. Therefore, I recommend acceptance of this paper.