List of Papers Browse by Subject Areas Author List
Abstract
It is challenging to discriminate autism spectrum disorder (ASD) from a highly heterogeneous database, because there is a great deal of uncontrollable variability in the data from different sites.
Recently, prompt learning has received considerable attention in domain adaptation as a promising solution.
However, its application to graph data like multi-site brain networks has not been fully studied.
It faces two major challenges: (1) complex graph structure; and (2) inter-individual variability.
To overcome the issues, we propose a novel prompt-tuning paradigm for multi-site brain network analysis (BrainPrompt) using functional magnetic resonance imaging (fMRI).
Specifically, we introduce two tunable soft prompts:
(1) a mask prompt to prune noisy edges while preserving important connections, and distill it to reduce domain-specific biases;
(2) sample prompts to capture inter-individual variations.
Our model outperforms other models on the ABIDE dataset, especially at sites with limited samples (e.g., the Stanford site, which has only 39 samples).
BrainPrompt achieves a 35.88% improvement in accuracy compared to the State-of-the-Art (SOTA) method, highlighting its superiority in small sites.
Furthermore, our results demonstrate the interpretability and generalization of the proposed method.
Our code is available at https://github.com/zliuzeng/BrainPrompt.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1562_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zliuzeng/BrainPrompt
Link to the Dataset(s)
The Autism Brain Imaging Data Exchange (ABIDE) dataset: http://fcon_1000.projects.nitrc.org/indi/abide/
BibTex
@InProceedings{ZhaLiu_BrainPrompt_MICCAI2025,
author = { Zhang, Liuzeng AND Li, Lanting AND Cao, Peng AND Yang, Jinzhu AND Zaiane, Osmar R.},
title = { { BrainPrompt: Domain Adaptation with Prompt Learning for Multi-site Brain Network Analysis } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {161 -- 171}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper explores domain adaptation using prompt learning to mitigate site factors and inter-individual variability in fMRI data acquired from different scanners. The proposed model achieves superior performance with lower computational cost, demonstrating its effectiveness even in small sample size sites with around 30 subjects. Furthermore, beyond reducing domain-specific biases, the approach leverages sample prompts to identify inter-individual variation.
- 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) By employing a domain adaptation approach, the model achieves superior ASD diagnosis performance even on datasets from unseen sites, outperforming other models.
2) The method attempts to reduce site heterogeneity and individual variability by leveraging domain-domain similarity and sample-domain similarity scores.
- 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 introduction claims that “Brain-Prompt achieves superior performance with lower costs,” but there is no clear evidence of reduced computational cost or time efficiency (e.g., model size or FLOPs).
2) Figure 4 highlights how the mask prompt identifies ASD key areas, but demonstrating effective domain adaptation is more critical. Reporting domain-domain similarity or sample-domain similarity scores would better align with the paper’s contributions, yet these are not discussed.
3) Section 2.2 on source domain adaptation lacks clarity in notation. The meaning of MMM in the M×M mask prompt is unclear—does it refer to functional connectivity? Similarly, the significance of D in domain-specific embedding is not well explained.
- Please rate the clarity and organization of this paper
Poor
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The primary criterion for scoring is whether the prompts effectively capture domain similarity and individual similarity to reduce differences. While the classification performance has improved, there is a lack of clear interpretation to confirm that this improvement is specifically due to domain adaptation.
- 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.
Thank you for answering all of my questions. I believe my requests align with the main contribution, and you have responded thoroughly.
Review #2
- Please describe the contribution of the paper
This paper introduces BrainPrompt, a prompt-tuning framework for domain adaptation in multi-site brain network analysis. The method addresses challenges posed by heterogeneous multi-site fMRI data and individual variability in autism spectrum disorder (ASD) diagnosis. Key contributions include 1) A mask prompt to prune noisy edges and preserve discriminative connections in brain networks; 2) Sample prompts to capture inter-individual variability; and 3) Knowledge distillation to reduce domain-specific biases and enhance cross-domain knowledge transfer. The framework is evaluated on the ABIDE dataset, demonstrating superior performance over state-of-the-art (SOTA) methods, particularly in sites with limited samples.
- 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 prompt learning to multi-site brain network analysis offers a n interesting perspective on domain adaptation for graph-structured data.
- The pepper is well-organiged and easy to fellow.
- 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 does not provide sufficient detail on how the mask prompt is specifically implemented. While it mentions that the mask prompt is used to prune noisy edges and preserve important connections, the exact formulation and derivation of the mask prompt are unclear. This lack of detail makes it difficult to assess the novelty and effectiveness of the approach.
- In my understanding, the mask prompt is similar to pseudo labels in domain adaptation, therefore, what is the domain-invariant prompt?
- 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.
- 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.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The prompt learning for multi-site brain network analysis is interesting but the details in the implementation are lacked.
- 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.
All of my concerns are well addressed.
Review #3
- Please describe the contribution of the paper
This paper presents a novel prompt learning strategy for multi-site brain network analysis, utilizing mask prompts and sample prompts to explore the topological structure of brain networks, even amidst individual heterogeneity. The experimental results demonstrate the advantages of the proposed method, particularly at sites with limited samples, underscoring its clinical significance.
- 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 present the first attempt to apply the prompt learning to the field of multi-site brain network analysis. The methodology is well-motivated to deal with the sparse and potentially noisy edges and significant individual variability.
- The prompts are developed at two levels: mask and sample. The mask prompts are used to prune noisy edges with domain-specific biases, while the sample prompts capture individual variability. This strategy is straightforward and effectively addresses the specific challenges in brain network analysis.
- The target prompt is initialized effectively by distilling knowledge from multiple source domains, which benefits to advance cross-domain knowledge transfer.
- 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 prompts are applied separately to masks and samples to explore intra-graph edge properties and inter-graph individual information. The authors are encouraged to investigate the relationship between intra- and inter-graph information, such as transferring insights from mask prompts to sample prompts.
- The effectiveness of the proposed method heavily depends on the pre-training stage and initial prompts. The authors should validate the method in more challenging scenarios where target domains exhibit a significant domain gap from the source domain, which could lead to instability in model initialization.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
This work is of high overall quality, but there are aspects of the methodology that could be enhanced.
- Reviewer confidence
Somewhat confident (2)
- [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.
I would like to thank the authors for preparing the rebuttal. Most of my concerns are addressed. Therefore, I maintain my rating for this paper as “Accept”.
Author Feedback
We thank the reviewers (R1, R2, R3) for their positive comments on the strong results.We clarify the main points:
Positive Suggestions(R1) We independently evaluated the effects of sample prompts and mask prompt.Future work will focus on investigating the interaction between sample and mask prompts and improving robustness across more diverse domains.
Mask prompt(R2) 1) The mask prompt is a learned matrix,optimized in an end-to-end manner to capture critical connectivity patterns in the brain network.Specifically, the mask prompt is randomly initialized, with its values constrained to the range [0,1] using a Sigmoid function. It is then multiplied with the original adjacency matrix.We apply L1 regularization on the mask prompt to encourage its sparsity for emphasizing disease-relevant patterns.The optimal mask prompt P_mask is obtained as follows:argmin_(theta)( L_ce( h_(theta)X,Y) + lambda |P_mask|1,where L_ce is the cross-entropy, and h(theta) is the prediction of our model with theta involving the learnable parameters of P_mask. 2) Due to multi-center data discrepancies, each center’s mask prompt may exhibit domain-specific biases.Our prompt-based efficient transfer method assigns higher initialization weights to more similar domains,thereby highlighting common structures,which are considered as domain-invariant,and suppressing domain-specific noisy structures, which are considered as domain-specific.Furthermore,the statistical analysis shows that 84.23% of the top 30 disease-related edges are consistent across domains, indicating that disease-related structures are largely shared despite data discrepancies.Although the mask prompt is not domain-invariant, this finding supports its interpretability by revealing shared structures across domains.
Prompts interpretation(R3) Due to space limitations, many experimental results are not included. 1) We have investigated the domain shift among the sites in the embedding distance. Specifically, we compute the cosine distances among sites with the original embedding generated by baseline model (transformer-encoder) and the embedding generated by our model. Results show a 66.67% decrease wrt average distance (from 0.27 to 0.09). Notably, the distance from NYU to its farthest site, UM, decreases by nearly 65.71% (from 0.35 to 0.12). This suggests that our prompt-based transfer method provides more appropriate initialization to the target domain, thereby helping to reduce domain discrepancies. 2) We have analyzed the domain-domain similarity by taking NYU as the target domain. The range of similarity scores is 0.69–0.85. The higher similarities of NYU–UCLA (0.85) and NYU–MaxMun (0.83), indicate being consistent with prior findings of cross-site similarity in ASD studies via the comparable method, i.e. BrainDAS, which is an explicit sample alignment method[MIA BrainDAS, arXiv HMMD]. We analyze sample-domain similarity by computing and normalizing the average similarity of all NYU samples to each source domain.The range of similarity scores is 0.54–0.82.The highest average scores are for NYU–UCLA (0.79) and NYU–MaxMun (0.82), consistent with the domain-domain similarity scores. 3) We have conducted ablation studies by removing either mask prompt or sample prompts (see Fig.3). The results also show that our prompt-based method helps mitigate domain shift. We can add interpretation the learned prompts in the camera-ready.
Cost comparison(R3) For transfer learning, it is crucial to evaluate and compare the number of parameters to be updated in the target domain.Our method updates only the prompt(1.95M parameters), which is 80.52% fewer than native fine-tuning method that updates all parameters(10.01M) in the target domain. Compared to other domain adaptation methods (e.g., maLRR: 8.09M; LRCDR: 12.51M), our model achieves competitive efficiency.
Confusing descriptions(R3) M is the number of ROIs in functional connectivity, and D is the feature dimension. It is set to M×(M−1)/2.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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”.
While some of the reviews are positive, it’s worth considering that two of them suggest the option for a rebuttal. Furthermore, reading the paper and the feedback from al reviewers, I think the paper lacks clarity on some of the choices and explanations on how the proposed approach actually addresses domain adaptation. It is true that the results show improvement against other methods, but whether the prompts and embeddings actually capture domain information is not explored.
Considering these limitations and the reviewer’s feedback, I would like to invite the authors for rebuttal to address the concerns raised by the reviewers and to further clarify the paper.
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A