List of Papers Browse by Subject Areas Author List
Abstract
Brain imaging classification is commonly approached from two perspectives: modeling the full image volume to capture global anatomical context, or constructing ROI-based graphs to encode localized and topological interactions. Although both representations have demonstrated independent efficacy, their relative contributions and potential complementarity remain insufficiently understood. Existing fusion approaches are typically task-specific and do not enable controlled evaluation of each representation under consistent training settings. To address this gap, we propose a unified cross-view contrastive framework for joint imaging–ROI representation learning. Our method learns subject-level global (imaging) and local (ROI-graph) embeddings and aligns them in a shared latent space using a bidirectional contrastive objective, encouraging representations from the same subject to converge while separating those from different subjects. This alignment produces comparable embeddings suitable for downstream fusion and enables systematic evaluation of imaging-only, ROI-only, and joint configurations within a unified training protocol. Extensive experiments on the ADHD-200 and ABIDE datasets demonstrate that joint learning consistently improves classification performance over either branch alone across multiple backbone choices. Moreover, interpretability analyses reveal that imaging-based and ROI-based branches emphasize distinct yet complementary discriminative patterns, explaining the observed performance gains. These findings provide principled evidence that explicitly integrating global volumetric and ROI-level representations is a promising direction for neuroimaging-based brain disorder classification. The source code is available at https://github.com/vveii14/circa-brain.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6144_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/vveii14/circa-brain
Link to the Dataset(s)
ADHD-200 dataset: http://preprocessed-connectomes-project.org/adhd200/
ABIDE dataset: http://preprocessed-connectomes-project.org/abide/
BibTex
@InProceedings{LiaWei_Joint_MICCAI2026,
author = { Liang, Wei AND He, Lifang},
title = { { Joint Imaging–ROI Representation Learning via Cross-View Contrastive Alignment for Brain Disorder Classification } },
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 paper proposes a unified cross-view contrastive framework for joint modeling of global volumetric imaging and ROI-based graph representations. By aligning these heterogeneous representations in a shared latent space, the method enables consistent and controlled comparison of imaging-only, ROI-only, and joint configurations under identical training settings. Empirical results on ADHD-200 and ABIDE suggest that the proposed alignment facilitates improved classification performance and captures complementary information across representations.
- 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 has several strengths. First, it proposes a unified framework that integrates global imaging and ROI-based graph representations through cross-view contrastive alignment, which is a reasonable and well-motivated approach beyond simple fusion methods.
Second, the study provides a controlled comparison of imaging-only, ROI-only, and joint models under the same training setting, which is valuable for understanding their relative contributions.
Third, the method is evaluated on two well-known datasets (ADHD-200 and ABIDE) and shows consistent, albeit modest, performance improvements across different model backbones. The large sample size is also a strength.
- 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 has several notable weaknesses.
First, both the imaging and ROI-based graph representations are derived from the same underlying data, raising concerns about whether the two views provide truly complementary information or largely redundant representations. This limits the strength of the multi-view learning claim.
Second, the novelty of the method is somewhat limited, as cross-view contrastive learning and multi-modal alignment have been widely studied in prior work (e. g. , contrastive representation learning frameworks such as SimCLR, MoCo, and CLIP). The paper does not clearly demonstrate what fundamentally new methodological contribution is introduced beyond applying these ideas to this specific setting.
Third, the reported performance improvements are relatively modest (e. g. , around 1–2% AUC, ADHD 72.73 vs 70.92, ABIDE 64.08 vs 62.84), making it unclear whether the gains are meaningful.
Fourth, given the use of subject-level contrastive alignment, it is important to clarify whether strict subject-level data splitting was enforced. Without this, there is a risk of subject-level leakage, especially in multi-site datasets such as ADHD-200 and ABIDE (the datasets are known for significant heterogeneity and is highly noisy).
Finally, although interpretability analyses are included, the paper does not clearly identify which specific features or brain regions drive the classification performance, limiting the neuroscientific insight 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 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.
(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 proposes a reasonable and well-motivated framework for integrating global imaging and ROI-based graph representations via cross-view contrastive learning. The idea is intuitive and the experimental setup is generally sound, with evaluation on two widely used neuroimaging datasets.
However, several factors limit the overall impact of the work. First, the methodological novelty is relatively limited, as the framework mainly adapts existing contrastive learning and multi-view alignment techniques to this setting without introducing fundamentally new components. Second, the reported performance improvements are modest (on the order of 1–2% AUC), and no statistical significance analysis is provided, making it unclear whether the gains are meaningful.
Third, there are concerns regarding the independence of the two representations, since both the imaging and ROI-based graphs are derived from the same underlying data, which may limit the claimed complementarity. In addition, the use of subject-level contrastive alignment raises the possibility of subject-level leakage, which is not clearly addressed in the paper.
Finally, while interpretability analysis is included, it remains relatively high-level and does not clearly identify specific brain regions or features driving the classification performance, limiting the neuroscientific insight.
Overall, while the paper is technically sound and potentially useful, these limitations place it slightly below the acceptance threshold in its current form.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 rebuttal clarifies several implementation details, particularly regarding subject-level splitting, which alleviates concerns about potential subject-level leakage during cross-validation. The authors also provide additional clarification regarding the motivation of the unified framework and the role of contrastive alignment.
However, my main concerns remain insufficiently addressed. In particular, the manuscript still provides limited evidence that the imaging and ROI-based representations capture truly complementary rather than partially redundant information, given that both views are derived from the same underlying imaging data. While the joint model shows consistent improvements, the reported gains remain relatively modest, and the claimed statistical significance is not quantitatively substantiated in the manuscript or rebuttal (e.g., no reported p-values, confidence intervals, or detailed statistical test results).
In addition, the evaluation relies entirely on internal 5-fold cross-validation within the same datasets, without external validation or independent cohort testing. Given the substantial heterogeneity and site variability in datasets such as ADHD-200 and ABIDE, the lack of external validation limits the strength of the generalizability claims.
Finally, the methodological contribution remains incremental relative to existing cross-view contrastive learning and multi-view alignment frameworks. Overall, while the work is technically sound and the experiments are reasonably comprehensive, I do not believe the rebuttal sufficiently resolves the concerns regarding novelty, complementarity, and generalizability.
Review #2
- Please describe the contribution of the paper
This paper presents a method for learning representations of structural brain imaging through contrastive alignment of the imaging and ROI data. The proposed method was tested on 2 public datasets with structural brain MRI for 2 brain disorder classification tasks (ADHD, ASD). The proposed approach performed better than imaging only or ROI only methods, with experiments including testing different encoders for the imaging and ROI data.
- 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 addresses the challenge of different types of ways to input structural brain data - using the original imaging volume vs the popular ROI-based methods. Considering both representation methods jointly allows for leveraging advantages of differing representations.
2.The proposed cross-view contrastive alignment is an interesting approach for the representation learning. The idea is that the learned embedding for the imaging and ROI information for the fMRI sample should be aligned in the learned representation space.
3.The experiments are performed on 2 public datasets and compared multiple different encoder backbones, under five-fold cross-validation. This helped show the generalizability of the proposed approach.
4.The ablation study on different ways to combine or fuse information from the two imaging and ROI representations helped to show the contribution of the proposed contrastive learning approach.
5.The experiments included a study to see the effect of removing a single branch input (imaging or ROI), simulating the case of “missing” data. The bounded drop as the amount of missing data was increased helped to verify that that the cross-view representation was effective in learning robust representations.
- 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.Missing ablation on loss contribution The method has a lambda term to weight the loss of the the contrastive learning vs brain disorder classification terms. The setting for lambda is not included. In addition, an ablation varying lambda would be informative to understand the loss contribution (I believe the case for no contrastive loss was included in the experiments for the “simple concatenation” setting.)
2.No statistical testing for differences in model results. I appreciate the reported variance across five-fold cross-validation. Additional statistical testing to show which methods / model settings are significantly different (i.e. in Table 1 and 2) would give stronger evidence for the better performance of the proposed approach.
3.Incorrect reference Ref. [25], which should be for the ABIDE dataset, is incorrect. It should cite Di Martino et al., “The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism,” 2014.Please check and correct this.
- 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
-
The authors state that “stratified” cross-validation was performed - could you clarify how the data are stratified? (ie., by diagnosis label, by site?)
-
Additional details on data preprocessing would be ideal to include.
-
In Fig. 1, I believe the dotted blue line from the projected encoder representation to the contrastive alignment module is not drawn from the proper location - should it not come from the p_img block (analogous to the line from the p_roi block)? Currently it is drawn as coming from z_img, before p_img.
-
- 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?
This work proposes an interest approach of cross-view contrastive learning of imaging and ROI-based representations of structural brain data. The experiments under different datasets, classification tasks, and baseline encoder settings and ablation studies on different ways to combine representations show the promise of the proposed contrastive method. Some additional details and empirical analysis would further strengthen the paper.
- 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.
This paper presents an interesting adaptation of contrastive learning to learn from multi-view brain MRI. The authors clearly addressed my questions and concerns in the rebuttal. I also found their response to other reviewers’ concerns reasonable, and important concerns brought up regarding methodology/experimental settings were addressed. While I agree with other reviewers that empirical gains are small, they are consistent in experiments. While I acknowledge others’ concerns regarding novelty, I still believe this application of contrastive learning to cross-view analysis of brain MRI is a unique and new technical contribution. Therefore, my final rating of this paper is accept.
Review #3
- Please describe the contribution of the paper
The paper proposes a cross-view contrastive learning framework that jointly trains global (volumetric imaging) and local (ROI-graph) representations for brain disorder classification, evaluated on ADHD-200 and ABIDE datasets.
- 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 proposes to leverage the complementary nature of image and graph representations for improving the performance.
The proposed fusion protocol provides improvement in performance compared to imaging-only or ROI-only configurations as shown in ablation studies.
An attempt is also made to incorporate interpretability into the analysis—showing that different configurations capture distinctive patterns for decision-making.
- 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.
- As cross-view or cross-modal contrastive learning is a well-established paradigm, using bidirectional subject-level contrastive objectives to brain imaging is not conceptually very novel.
- The statement—Existing fusion approaches are typically task-specific—is neither clarified nor comparatively analyzed against them.
- While the overview of the proposed framework indicates that diverse imaging modalities can be used, all the experiments are based only on the anatomical MRI datasets.
- Abstract mentions a bidirectional contrastive objective that encourages representations from the same subject to converge while separating those from different subjects. This is standard instance-discrimination contrastive loss. It’s unclear what “bidirectional” adds—whether the loss operates symmetrically across both branches or introduces some asymmetric cross-view alignment.
- To assess robustness under missing views, the experiments are conducted to simulate view absence by zero-masking one branch for a random subset of subjects. However, the only case that is realistic is the image being available and the ROI missing due to preprocessing failure. ROI availability and missing images don’t make sense; ROI-based features (parcellations, connectivity matrices, and graph nodes) are derived from the same underlying brain image.
- Lastly, comparison with the state of the art is not quite fair or comprehensive. The significance of the proposed InfoNCE loss should have been demonstrated in comparison with other contrastive losses (like triplet loss and decoupled contrastive loss) and other non-contrastive losses.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Novelty is incremental at best Lack of a comprehensive comparison with state-of-the-art
- 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 and AC for their constructive feedback and clarify key points below.
Review 1 & Review 3 – Methodological novelty We agree that our method builds on established contrastive learning. Our contribution is not a new loss, but a unified and controlled framework for jointly modeling volumetric imaging and ROI-graph representations. Prior fusion approaches are often architecture-specific and lack systematic analysis. In this work, we provide: (i) cross-view alignment between global volumetric and ROI-level graph features, (ii) controlled comparisons (imaging-only, ROI-only, joint) across multiple backbones, and (iii) interpretability analyses revealing complementary neuroanatomical patterns. This establishes a principled framework for imaging–ROI complementary learning.
Review 1 – Performance & complementarity Although absolute gains over imaging-only and ROI-only baselines are modest, the improvements are consistent across datasets, encoder choices, and evaluation metrics. As shown in Table 1, we observe dataset-specific trends: imaging-based models perform better on ADHD, whereas ROI/graph-based models perform better on ABIDE. Despite these differences, the proposed joint framework consistently achieves the best overall performance on both datasets, further supporting the complementary nature of the two representations rather than simple redundancy.
Review 1 & Review 2 – Statistical significance We conducted paired t-tests on (subject-level) Accuracy and DeLong’s tests on AUC, both confirming statistical significance. We will include these results in the revision.
Review 1 – Subject-level leakage in contrastive alignment Strict subject-level splitting was enforced in 5-fold cross-validation, with each subject assigned exclusively to either the training or test set. Contrastive alignment was performed only within the training set, and test subjects were never involved in optimization. Therefore, subject-level leakage is strictly prevented.
Review 1 – Interpretability Our interpretability analysis aims to assess whether the two branches capture complementary patterns. We agree that deeper region- and feature-level interpretation would strengthen the neuroscientific insight. The framework is backbone-agnostic and can be combined with more advanced interpretable imaging and graph models in future work.
Review 2 – Hyperparameters (λ) Hyperparameters were optimized via Optuna. The loss-weight λ was searched in [0.05, 0.5], with optimal values 0.48 (ADHD) and 0.2 (ABIDE), indicating moderate alignment strength. The “Concat” setting in Table 2 corresponds to λ = 0, serving as an implicit ablation of the contrastive objective. We will clarify it in the revision.
Review 2 – Clarifications 1.We will update the ABIDE reference (Ref. 25) to Di Martino, A. et al., 2014.2.Stratification is diagnosis-label based at the subject level. 3.The dotted blue line should indeed originate from p_img (not z_img); we will update Fig. 1 accordingly.
Review 3 – Additional clarifications 1.“Task-specific fusion” refers to methods tied to specific architectures or pipelines, where performance gains may reflect model complexity rather than representation quality. In contrast, our approach focuses on studying the complementary contributions of imaging and ROI feature representations under a unified setting. 2.“Bidirectional” refers to the symmetric matching formulation in Eq. (5), where contrastive matching is computed in both directions between the imaging and ROI branches. The two directional losses are averaged over the mini-batch, and no additional asymmetric alignment component is introduced.
Review 3 – Loss comparison We agree that broader loss comparisons would be valuable. However, our goal is not exhaustive benchmarking but validating a unified framework under controlled settings. InfoNCE is used due to its suitability for subject-level pairing. We will clarify this scope and discuss extensions as future work.
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 paper addresses a relevant and timely problem with a reasonable and technically sound framework, and reviewers acknowledge its promise across datasets and experimental settings. However, concerns remain about limited methodological novelty, modest and potentially insignificant performance gains, unclear independence between representations, and possible subject-level leakage. The rebuttal should specifically clarify: (1) statistical significance of the reported improvements, (2) safeguards against subject-level leakage, and (3) the extent of true complementarity between views.
- 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.
This paper presents a technically sound and interesting application of cross-view contrastive learning to brain MRI, with consistent though modest empirical gains. Given the generally positive consensus among reviewers and the rebuttal’s clarification of key methodological concerns, I recommend 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.
Reject
- Please justify your recommendation.
After rebuttal, R1 changed from Weak Reject to Reject, while R2 maintained Accept and R3 remained Weak Accept. Although authors clarified concerns regarding subject-level leakage, statistical testing, and the motivation for cross-view alignment, concerns remained about the limited methodological novelty, modest performance gains, lack of external validation, and insufficient evidence that image and ROI representations provide truly complementary rather than redundant information. Given these unresolved issues, I recommend Reject.
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.
The paper is technically sound and provides a controlled framework for jointly modeling volumetric imaging and ROI-graph representations. Although the novelty is incremental and the gains are modest, the rebuttal addresses the key concerns about subject-level leakage, statistical significance, loss weighting, and bidirectional alignment. The final version should tone down broad novelty claims and clarify the limits of the complementarity analysis.
