List of Papers Browse by Subject Areas Author List
Abstract
Zero-shot learning (ZSL) is critical for deep learning models being deployed in unseen downstream applications. Given that fMRI studies of the human connectome with respect to cognitive disorders are boutique and lack sufficient labeled samples, a reliable and interpretable ZSL technology is necessary to empower the brain foundation model for clinical applications. Although self-supervised learning and transfer learning on data reconstruction and semantic information, respectively, have achieved success in ZSL performance for language and vision, little attention has been paid to the recognition of brain disordering. In contrast to stereotypical language or vision data, the human brain is a dynamically wired system where distributed regions communicate through functional connectivity and spontaneously respond to stimuli from environmental exposures. Thus, functional neuroimages are often associated with phenotypic traits underlying brain-environment interactions (BEIs), such as cognitive states and clinical outcomes. By capitalizing on large-scale functional neuroimages as well as a rich collection of BEI data, we break the frame of self-supervised and transfer learning by using logical regression as the pre-training objective for brain connectome. We formulate ZSL on unseen classes by identifying a reliable matching across environmental variables, which is derived from a decoder-only model for BEI prediction from functional connectivity. Together, we present a novel learning schema of brain-environment cross-attention (BECA) meta-matching, which is a new horizon of ZSL for brain connectome. In experiments, all fMRI data in HCP-young adult and HCP-aging datasets are utilized for pre-training, and BECA is evaluated on disease early diagnosis of Autism, Parkinson’s disease, and Schizophrenia, where promising results indicate the great potential to facilitate current neuroimaging applications in clinical routines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2086_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Chrisa142857/brain_network_decoder/tree/zero-shot-learning
Link to the Dataset(s)
HCPA dataset: https://www.humanconnectome.org/
HCPYA dataset: https://www.humanconnectome.org/
ADNI dataset: https://adni.loni.usc.edu/
PPMI dataset: https://auckland.figshare.com/articles/dataset/NeurIPS_2022_Datasets/21397377
ABIDE dataset: https://auckland.figshare.com/articles/dataset/NeurIPS_2022_Datasets/21397377
BibTex
@InProceedings{WeiZiq_BrainEnvironment_MICCAI2025,
author = { Wei, Ziquan AND Dan, Tingting AND Wu, Guorong},
title = { { Brain-Environment Cross-Attention (BECA) Meta-Matching: A New Perspective of Brain Connectome Zero-Shot Learning } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {140 -- 149}
}
Reviews
Review #1
- Please describe the contribution of the paper
A brain-environment cross-attention based meta-matching method was proposed for brain disease classification using functional connectivity data, under the zero-shot and few-shot learning setting.
- 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.
It is interesting and novel to learn the cross-attention between FC and Environment information.
- 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.
- Lack of detailed information about the method, which makes it difficult to follow. Please see comments below.
- The performance was not good under the meta-matching setting.
- 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.
- 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/2025/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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- It is not clear how the LCM was pretrained. For example, how many tasks (BEIs) were used, what losses were used, how to balance losses for brain state (for HCP?) and brain disease (for ADNI?) prediction.
- It is not clear how the cross-attention meta-matching works under the ZSL setting. How was the BECA maps calculated for the unseen subjects?
- The derivation of Eq.(3) was not intuitive. More explanation would be helpful.
- In Fig.3, the “white-matter surface” is confusing (it seems to be a pial surface actually). Does it refer to an empty BECA map here?
- The number of parameters in the proposed method (1.2B) was way more than other models under comparison (<=30M). How does the model size affect the performance?
- It seems that the ZSL/FSL performance was not good under the meta-match setting. The AUC of the BECA (MM) model was only around or less than 0.5 across datasets (ABIDE, PPMI, AND SZ) with different finetuning data sizes (FT%).
- In addition to classification, can the proposed method be applied to regression tasks (e.g., prediction of brain age, cognition, etc.)?
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Review #2
- Please describe the contribution of the paper
BECA Framework: Introduces a novel zero-shot learning (ZSL) framework for brain connectome analysis using brain-environment interaction (BEI) cross-attention and statistical meta-matching. Large Connectome Model (LCM): Pre-trained on multi-dataset fMRI data (HCP/ADNI) via a two-stage strategy (momentum learning + semi-supervision), achieving 1.2B parameters. Clinical Validation: Demonstrates superior ZSL performance on Autism (ABIDE), Parkinson’s (PPMI), and Schizophrenia (SZ) datasets, with zero-shot F1 scores outperforming SOTA (e.g., 39.70 vs. 33.92 on ABIDE). Interpretability: BECA activation maps align with neuroscientific findings (e.g., default mode network in Autism), enhancing clinical plausibility. Open Source: Releases anonymized code and pre-trained weights for reproducibility.
- 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.
Innovation: First to integrate BEI-driven cross-attention for brain connectome ZSL, breaking from traditional self-supervised/transfer learning paradigms. Clinical Impact: Addresses data scarcity in early disease diagnosis (e.g., Parkinson’s with 128 samples) through meta-matching and few-shot fine-tuning. Rigor: Comprehensive evaluation across 3 datasets with 10-fold cross-validation, outperforming BrainMass, BolT, and NeuroPath in most scenarios. Transparency: Visualized BECA maps provide mechanistic insights into disease-related neural circuits.
- 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.
Baseline Limitations: Omits comparison with recent ZSL benchmarks (e.g., NeuroGraph [20]), potentially understating novelty. Computational Cost: LCM’s 32-layer architecture (1.2B parameters) raises scalability concerns, yet hardware requirements and training efficiency are unaddressed. Data Heterogeneity: Pre-training mixes healthy (HCP) and patient (ADNI) data; unclear how this impacts disease-specific ZSL generalization. Statistical Validity: Meta-matching relies on t-tests without multiple comparison correction, risking inflated false positives.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Compare with NeuroGraph [20] or other recent ZSL baselines. Discuss BECA maps’ correlation with clinical metrics (e.g., disease severity). Explore LCM architectural simplification (e.g., fewer layers) for clinical deployment.
- 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?
BECA’s methodological novelty (BEI-driven ZSL) and robust validation on challenging clinical datasets justify acceptance. Weaknesses are addressable via revisions.
- 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
Review #3
- Please describe the contribution of the paper
This article proposes BECA, an innovative zero-shot learning framework for early diagnosis of brain disorders. The framework effectively leverages brain functional connectivity and brain-environment interaction data, achieving reliable identification of unseen categories through self-attention and cross-attention mechanisms. Experimental results across multiple brain disorder datasets demonstrate BECA’s superior performance in both zero-shot and few-shot learning tasks, with high accuracy and interpretability. Methodologically, this work addresses the limitations of meta-matching approaches in enabling zero-shot learning, showing significant value. However, several questions remain to be addressed before considering publication.
- 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) This article introduces a novel ZSL framework termed “BECA Meta-Matching,” designed to advance connectomics through large-scale functional neuroimaging data and BEI data, ultimately supporting early diagnosis of brain disorders. 2) The BECA framework pioneers a reliable ZSL methodology for brain connectomics by leveraging logistic regression-based pretraining of connectome models and brain-environment interaction data. It transcends the limitations of conventional self-supervised learning and transfer learning approaches, offering fresh insights for neuropsychiatric diagnostics. This innovation demonstrates significant novelty in the field of connectomics. 3) Using multi-cohort datasets (e.g., HCP-Young Adult, HCP-Aging, ADNI) for pretraining, the authors evaluated BECA’s performance through zero-shot and few-shot learning assessments on ABIDE, PPMI, and schizophrenia datasets. These datasets encompass diverse brain disorders—including autism spectrum disorder, Parkinson’s disease, and schizophrenia—ensuring strong clinical relevance.
- 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) What constitutes BEIs and how are they related to phenotypes? Which specific BEIs were selected across different datasets? Clarify the selection criteria and processing methods for BEIs, including fMRI preprocessing parameters (e.g., number of nodes in AAL atlas). This is crucial for phenotype prediction work. Examples: [1] He T, et al. Meta-matching as a simple framework to translate phenotypic predictive models from big to small data. Nature Neuroscience, 2022. [2] He Z, et al. F2tnet: fMRI to T1w MRI knowledge transfer network for brain multi-phenotype prediction. MICCAI, 2024.
2) Does ZSL in this work involve single-dataset input for single-disease classification? If so, can the framework simultaneously infer across multiple disease datasets to determine individual disease? While enabling ZSL via meta-matching represents progress, the authors need to clarify whether multi-disease joint inference is supported under ZSL/FSL settings.
3) The author set τ = 0.05. What is the influence of τ on the experimental results under different Settings?
4) The manuscript lacks sufficient references to foundational work. Key additions needed: [1] Zero-shot and few-shot learning (ZSL and FSL) have demonstrated remarkable success in various domains… [2] tremendous efforts have been made to pre-train large models… [3] Meanwhile, in the recognition of brain disorders, where labeled samples are often limited and expensive to obtain… [4] brain fMRI is often associated with nonimaging phenotypes regarding brain-environment interactions…
Minor:
1) Define abbreviations at first mention (e.g., “functional magnetic resonance imaging (fMRI)”.
2) Address the incomplete brain illustration in Figure 1(c).
- 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.
- 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/2025/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?
The experiment design is persuasive enough to prove the effectiveness of this method. This work addresses the limitations of meta-matching approaches in enabling zero-shot learning, showing significant value.
- 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
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
Based on the two very good reviews from Reviewers #1 and #3, I recommend acceptance of this paper. The methodological contribution is of high interest, and I truly believe that the issues encountered by Reviewer #2 regarding understanding of the method can be corrected easily.