acceptodds
Under review as a conference paper at ICLR 2027

CBT-GCL: Leveraging Connectional Brain Templates for Enhanced ASD Diagnosis via Graph Contrastive Learning

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder with a rising global prevalence, making its early and precise diagnosis crucial. Although deep learning has shown great potential in the diagnosis of ASD with Functional Connectivity (FC) data, it still suffers from challenges, including substantial noise in raw FC data, redundant connections, and high inter-individual heterogeneity. In facing these challenges, the Connectional Brain Template (CBT), which represents a group of brain networks, can effectively mitigate them, thereby providing prior guidance for diagnosis. To this end, we propose a CBT-driven graph contrastive learning model, called CBT-GCL, which utilizes CBT as prior knowledge to assist ASD diagnosis. Specifically, our model concurrently trains two encoders to extract multi-level features from both the CBT and individual-specific FC data, subsequently fusing and enhancing these features. A perceptual loss function aligns hierarchical features between the two groups, while a contrastive loss function pulls features of positive samples closer and pushes features of negative samples apart. This effectively extracts critical connections from FC data, thereby enhancing ASD diagnostic performance. Experiments on the Autism Brain Imaging Data Exchange (ABIDE) I and II datasets demonstrate that CBT-GCL outperforms state-of-the-art models in multiple evaluation metrics. Ablation studies further confirm the effectiveness of each component of CBT-GCL.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain network, connectional brain template, functional brain network analysis, ASD diagnosis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZxYLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CBT-GCL uses population-level connectivity templates to regularize individual brain-graph representations for ASD classification. Two category-specific templates are constructed with the existing Deep Graph Normalizer using training subjects only. Both templates are added to every individual functional-connectivity matrix, and the fused connectivity profiles are also included in the node features.

During training, an individual-graph encoder and a template encoder produce multi-level representations. Semantic Feature Alignment encourages agreement between individual and label-matched template features. Contrastive Feature Enhancement combines their final representations into an augmented view and applies a supervised contrastive objective. The classifier uses the individual encoder’s low- and high-level features. At inference, the auxiliary template branch is removed, although both fixed templates remain part of the input.

Critical strengths
  1. The complete method improves classification results on both datasets.

  2. The module ablations support the usefulness of the combined training procedure.

Critical weaknesses
  1. The experiments do not adequately isolate the proposed template-guidance contribution. Table 3 compares the complete method with a baseline that has neither CBT guidance nor a contrastive objective. Equations 10–12 use diagnostic labels to attract all same-class representations, so some improvement could come from supervised contrastive regularization independently of the connectional templates. The paper does not report a matched contrastive model without CBT guidance. It also does not compare DGN templates with simpler category-wise connectivity templates or learned class prototypes. These controls matter because the proposed auxiliary branch receives only two fixed category templates, selected by the training label. The existing ablations show that the combined procedure helps, but do not establish why learned connectional templates are needed rather than a simpler class-conditioned regularizer. A direct comparison with the template-registration approach discussed as the main motivation is also absent.

  2. The evaluation is too limited to establish that the reported advantage is reliable. Section 4.1 uses a random 70/20/10 split of subjects pooled across sites. Given the dataset sizes, the test partitions contain approximately 110 and 75 subjects. Tables 2, 3, and 5 provide point estimates without reporting repeated splits, seed variability, confidence intervals, or paired comparisons. This is consequential because the AUC improvements over the strongest baselines are only 0.009 on ABIDE I and 0.003 on ABIDE II. I cannot determine whether these differences are stable across reasonable splits and training runs. The template-transfer experiment does not resolve this concern: Appendix G still trains the downstream classifier on ABIDE II and changes only the template source. It is useful evidence about template reuse, but not a test of classifier generalization to unseen sites or populations. The current evidence therefore supports a promising within-dataset result more clearly than the broader robustness and diagnostic claims.

Questions for authors
  1. Clarification of the existing evaluation: Were the reported numbers obtained from one split and one training run, or averaged over multiple runs? Please provide the exact test counts, split stratification, checkpoint-selection rule, and whether all baseline methods used identical subject partitions. If repeated results already exist, their variability would materially affect my assessment. Assessment-changing control: How does the method compare with the same FC encoder and supervised contrastive objective, but without CBT guidance? An input-fusion-only variant would also help distinguish the effects of adding templates to the graph from those of SFA and CFE.
  2. Assessment-changing control: Does DGN-based guidance outperform category-wise mean or median connectivity templates, or learned class prototypes, with the downstream architecture and losses held fixed? This comparison is important for assessing the specific contribution of the connectional-template construction.
  3. Clarification of validation scope: Is there any site-held-out evaluation beyond the pooled random splits? If not, please distinguish template-source transfer from classifier transfer in the conclusions. For any such evaluation, templates and all model-selection steps would need to exclude the held-out sites.
1 comment
AuthorsOpenReview ↗

Response to Reviewer ZxYL

Dear Reviewer,

Thank you for your constructive comments. We have addressed all the comments with specific explanations and supplemented the corresponding experiments. As the author rebuttal period is coming to an end, we would like to ask if your concerns have been resolved and if you have any further questions or concerns. We would be happy to answer any questions while the discussion remains open.

If our responses have resolved all your concerns, we would greatly appreciate it if you could let us know whether you would reconsider the score for this paper.

Thank you again for your time and valuable feedback.

Sincerely,

The Authors

Reviewer_81nYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the challenges of ASD dignosis and proposes a CBT-driven graph contrastive learning model, called CBT-GCL, which utilizes CBT as prior knowledge to assist ASD diagnosis. Specifically, the model concurrently trains two encoders to extract multi-level features from both the CBT and individual-specific FC data, subsequently fusing and enhancing these features. A perceptual loss function aligns hierarchical features between the two groups, while a contrastive loss function pulls features of positive samples closer and pushes features of negative samples apart. This effectively extracts critical connections from FC data, thereby enhancing ASD diagnostic performance. Experiments on the Autism Brain Imaging Data Exchange (ABIDE) I and II datasets demonstrate that CBT-GCL outperforms state-of-the-art models in multiple evaluation metrics. Ablation studies further confirm the effectiveness of each component of CBT-GCL.

Critical strengths

The paper is clear and easy to follow. The approach seems to be simple and effective.

Critical weaknesses
  1. There are some missing related work, such as Com-BrainTF [1], which also explore Connectional Brain Template and community information for improving ASD diagnosis. The authors should discuss the difference from it and compare their approach against it.
  2. The experimental results are not convincing. For example, brain network transformer[2] officially report their performance on ABIDE as 80.2%, while the authors report the performance with significantly lower value. The authors should clarify the underlying reason about it.

[1]Community-aware transformer for autism prediction in fmri connectome. MICCAI 2023

[2] Brain network transformer.Neurips2022.

Questions for authors

See above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 81nY

Dear Reviewer,

Thank you for your constructive comments. We have addressed all the comments with specific explanations and supplemented the corresponding experiments. As the author rebuttal period is coming to an end, we would like to ask if your concerns have been resolved and if you have any further questions or concerns. We would be happy to answer any questions while the discussion remains open.

If our responses have resolved all your concerns, we would greatly appreciate it if you could let us know whether you would reconsider the score for this paper.

Thank you again for your time and valuable feedback.

Sincerely,

The Authors

Reviewer_H5nEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CBT-GCL uses class-specific connectional brain templates to guide an individual-connectome encoder for ASD classification. Multi-level alignment and supervised contrastive enhancement transfer template information during training; inference uses both fixed templates without test labels. Evaluation covers ABIDE I and II.

Critical strengths
  1. Training-only template construction and label-free inference are explicitly specified, making the population-prior mechanism usable at deployment (Appendices B-E).
  2. Reported improvements on both datasets, component ablations, and template-transfer experiments provide preliminary evidence that population priors help (Tables 2, 3, 5).
Critical weaknesses
  1. The gains lack uncertainty estimates. The pooled-site 70/20/10 protocol reports neither repeated-split uncertainty nor held-out-site results; AUC gains over the best baseline are only 0.009 and 0.003 (Table 2). Appendix G transfers templates but retrains the classifier, leaving cross-dataset classifier generalization untested.
  2. The template contribution is not isolated. Table 3 lacks input-fusion-only, class-mean-template, and contrastive-without-CBT controls. The closely related template-registration approach discussed in Section 2 is also missing from Table 2.
  3. Signed connectivity is unresolved. Equation (18) generates nonnegative distance-based templates, whereas Equation (16) retains signed Pearson correlations. Adding these templates directly in Equation (22) can alter edge signs; neither this effect nor its implications for the claimed denoising mechanism is evaluated.
Questions for authors
  1. Do the gains persist across repeated splits and held-out sites, with paired confidence intervals?
  2. How much improvement remains over class-mean templates and contrastive learning without CBT under the same encoder?
  3. How are negative population connections represented, and how often does template addition change edge signs?
2 comments
AuthorsOpenReview ↗

Response to Reviewer H5nE

Dear Reviewer,

Thank you for your constructive comments. We have addressed all the comments with specific explanations and supplemented the corresponding experiments. As the author rebuttal period is coming to an end, we would like to ask if your concerns have been resolved and if you have any further questions or concerns. We would be happy to answer any questions while the discussion remains open.

If our responses have resolved all your concerns, we would greatly appreciate it if you could let us know whether you would reconsider the score for this paper.

Thank you again for your time and valuable feedback.

Sincerely,

The Authors

I appreciate the authors’ additional experiments and clarifications, particularly the targeted template controls, LOSO evaluation, and analysis of edge-sign changes. These additions substantially address my concerns about the template contribution and clarify the limitations of the connectivity interpretation. However, the reliability of the reported gains remains insufficiently established. The five-seed experiments do not clarify whether data splits were repeated, and the new evaluations lack matched baseline results and paired confidence intervals for the improvements. Thus, although the response provides useful evidence of generalization, it does not yet demonstrate consistent superiority over competing methods. I therefore maintain my original score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.