List of Papers Browse by Subject Areas Author List
Abstract
Modeling the complex interactions among functional subnetworks is crucial for the diagnosis of mental disorders and the identification of functional pathways. However, learning the interactions of the underlying subnetworks remains a significant challenge for existing Transformer-based methods due to the limited number of training samples. To address these challenges, we propose KD-Brain, a Prior-Informed Graph Learning framework for explicitly encoding prior knowledge to guide the learning process. Specifically, we design a Semantic-Conditioned Interaction mechanism that injects semantic priors into the attention query, explicitly navigating the subnetwork interactions based on their functional identities. Furthermore, we introduce a Pathology-Consistent Constraint, which regularizes the model optimization by aligning the learned interaction distributions with clinical priors. Additionally, KD-Brain leads to state-of-the-art performance on a wide range of disorder diagnosis tasks and identifies interpretable biomarkers consistent with psychiatric pathophysiology. Our code is available at https://github.com/LiuSy0731/KDBrain.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0073_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/LiuSy0731/KDBrain
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuSiy_Exploring_MICCAI2026,
author = { Liu, Siyu AND Wen, Guangqi AND Cao, Peng AND Yang, Jinzhu AND Liu, Xiaoli AND Wang, Fei AND Zaiane, Osmar R.},
title = { { Exploring Subnetwork Interactions in Heterogeneous Brain Network via Prior-Informed Graph Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes KD-Brain, a Prior-Informed Graph Learning framework for mental disorder diagnosis using functional MRI. To address the overfitting issue of pure data-driven Transformers on small-sample medical datasets, the authors introduce external prior knowledge from Large Language Models (LLMs). The core contributions include: (1) A Semantic-Conditioned Interaction mechanism that injects BioMedBERT-encoded disorder-specific semantic descriptions into the attention Query; (2) A Pathology-Consistent Constraint (PMC) that uses LLM-generated (e. g. , GPT-4) subnetwork interaction probabilities to regularize the learned attention distributions via KL-divergence. The method demonstrates state-of-the-art performance on ASD, BD, and MDD classification tasks.
- 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 idea of integrating LLM-derived clinical and semantic priors into Graph Transformers is highly innovative. Shifting from a blind statistical correlation fitting to a prior-informed learning paradigm is an insightful approach to mitigating the small-sample overfitting problem in neuroimaging. 2.The authors conducted extensive experiments across three mental disorders (ASD, BD, MDD) and compared their method against 12 competitive baselines (including recent Brain GNNs and Transformers), achieving leading classification accuracy and AUC. 3.By explicitly modeling subnetwork interactions, the framework uncovers functional pathways (e. g. , DMN->CEN->DMN for ASD) and critical brain regions (e. g. , Thalamus, Orbitofrontal Cortex) that align well with existing psychiatric and cognitive neuroscience literature, adding clinical value to the proposed AI model.
- 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.
Despite the clinical relevance of the problem and the promising empirical results, the current manuscript still has several important methodological and presentation-related weaknesses that should be addressed to strengthen the paper. 1.The use of disorder-specific priors raises concerns about potential prior leakage. A central methodological concern is that the proposed Pathology-Consistent Constraint (PMC) uses disorder-specific priors to regularize the learned attention distribution. In the context of a disease classification task, generating target distributions from disorder-specific textual descriptions may implicitly inject label information into the training objective. As a result, it is unclear whether the model is truly learning robust topological patterns from the fMRI data, or whether part of the performance gain comes from being guided by disease-specific prior information. The current ablation studies are not sufficient to rule out this concern. More stringent control experiments are needed, such as replacing the true priors with random priors, incorrect disease descriptions, or disorder-agnostic descriptions, to demonstrate that the improvement arises from the modeling framework itself rather than from task-specific prior injection. 2.The notion of “heterogeneous brain network” needs to be defined more carefully. The paper describes the brain network as a heterogeneous graph by partitioning 116 ROIs into three functional subnetworks (DMN, CEN, and SN). While this formulation does introduce functional grouping at the subnetwork level, it does not fully align with the standard definition of heterogeneous graphs in graph representation learning, which typically involves multiple node types, multiple edge/relation types, and relation-specific message passing mechanisms. In its current form, the method appears closer to a clustered or functionally partitioned homogeneous graph, rather than a rigorously defined heterogeneous graph model. Therefore, the terminology may be somewhat overstated and should be clarified more carefully to avoid a mismatch with established graph learning literature. 3.Limited evidence for the reliability and reproducibility of LLM-generated priors. The PMC module relies on GPT-4 to produce quantitative prior distributions over subnetwork interactions. However, the reliability of using LLMs as sources of precise probability distributions in a specialized clinical setting is not sufficiently justified. The paper does not provide a systematic sensitivity analysis with respect to prompt design, model version, or response variability, nor does it quantify how well these LLM-generated priors agree with expert annotations or established large-scale clinical evidence. Although the paper notes that GPT-4 and DeepSeek-R1 produce similar rankings, such evidence is still insufficient to establish the stability, validity, and reproducibility of these priors. 4.The stated motivation is not fully aligned with the actual technical contribution. One of the paper’s main motivations is that existing positional encodings fail to reflect functional brain organization. However, KD-Brain does not actually propose a new positional encoding scheme in the spatial or topological sense. Instead, it injects semantic conditions into the attention mechanism. This design may indeed improve semantic guidance during interaction modeling, but it does not directly address the positional encoding problem as framed in the introduction. As such, there is a noticeable gap between the stated motivation and the technical solution actually delivered.
- 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.
(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 presents an inspiring and creative approach to brain network analysis by leveraging LLM-generated clinical priors to guide Graph Transformers, with the goal of alleviating the small-sample challenge in fMRI studies. The empirical performance is promising, and the attempt to improve biological interpretability is commendable. My current rating is Weak Reject, mainly because the central methodological claim is not yet fully supported. In particular, the use of disorder-specific descriptions to formulate the PMC loss raises a substantial concern about potential prior leakage. In a disease classification setting, it remains unclear whether the reported gains come from improved topological representation learning on the brain graph, or from disease-specific information injected through the priors. The current ablation studies are helpful, but they are not sufficient to disentangle these two factors. Stronger control experiments, such as using random priors, disorder-agnostic priors, or mismatched disease descriptions, would be necessary to validate the claim more convincingly. In addition, the use of the term “heterogeneous graph” does not appear to be fully aligned with the standard graph learning definition, and the reliability of the LLM-generated quantitative priors is not yet sufficiently established. These issues do not negate the potential value of the work, but they do weaken its methodological clarity and theoretical grounding in the current form. Overall, I find the idea interesting and potentially impactful, but I do not think the present version provides enough evidence to fully support its main claims. I would be open to raising my score if the authors can convincingly address the prior leakage concern through stronger control experiments and clarify the graph formulation in the rebuttal.
- 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
5.This paper proposes KD-Brain, a prior-informed graph learning framework for modeling heterogeneous brain functional networks and diagnosing mental disorders. The core contributions are summarized as follows: (1) It addresses the critical limitations of Transformer-based brain network analysis (lack of functional semantics, overfitting to spurious correlations on small-scale medical fMRI data) by introducing dual prior knowledge integration (semantic prior + clinical pathological prior). (2) It designs the Semantic-Conditioned Interaction Learning (SSIL) module, which injects disorder-specific subnetwork semantic priors into the Transformer attention query to guide interpretable subnetwork interaction modeling. (3) It proposes the Pathology-Consistent Constraint (PMC) based on KL divergence, which regularizes the model by aligning learned interaction distributions with clinical prior knowledge to ensure neurobiological plausibility. (4) The framework achieves state-of-the-art performance on three mental disorder diagnosis tasks (ASD, BD, MDD) and identifies interpretable functional pathways and critical brain regions consistent with psychiatric pathophysiology.
- 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) Novel and targeted methodological design: The dual prior integration scheme directly solves the core pain points of pure data-driven graph Transformers in small-sample brain network analysis: lack of functional inductive bias and poor generalization. The semantic-conditioned attention and pathology-consistent constraint are well-motivated and tightly coupled with brain science knowledge. (2) Clinical and neuroscientific value: The model does not only pursue diagnostic accuracy but also outputs biologically interpretable biomarkers (disease-specific functional pathways, key brain regions), which align with established psychiatric pathophysiology and provide new insights for mental disorder mechanism research. (3) Comprehensive and rigorous experimental validation: The paper compares with 12 baselines, conducts systematic ablation studies (interaction order, module removal, attention replacement), and verifies the effectiveness of semantic priors and pathological constraints.
- 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) Incomplete performance analysis: The model achieves optimal performance on ASD and BD tasks but shows slight performance fluctuations on the MDD task. The paper does not analyze the underlying reasons for this phenomenon (e.g., data distribution, pathological complexity). (2) Single subnetwork partition setting: The experiment only uses three fixed functional subnetworks (DMN, CEN, SN). The robustness and generalization of the framework under other subnetwork partition strategies (e.g., different brain atlases, more fine-grained subnetworks) are not verified. (3) Insufficient analysis of LLM-generated priors: The paper uses GPT-4/DeepSeek-R1 to generate clinical prior distributions but does not compare the impact of different LLMs, manual clinical priors, or noisy priors on model performance, lacking robustness verification of the prior module. (4) Lack of hyperparameter and cross-center validation: The tuning range and sensitivity analysis of key hyperparameters (λ_sp, β) are not provided. The model is only validated on a single-center dataset and the NYU subset of ABIDE, without cross-center generalization experiments.
- 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
(1) Please supplement the analysis of the MDD task performance fluctuation and explain the difference in modeling effects among different mental disorders. (2) Add robustness experiments under different subnetwork partition strategies and brain atlases to verify the generalization of the framework. (3) Conduct ablation studies on LLM prior sources (e.g., different LLMs, manual clinical priors) to prove the stability of the prior module. (4) Add hyperparameter sensitivity analysis and cross-center validation to further improve the persuasiveness of the experiment.
- 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 paper puts forward a targeted solution for the problems of insufficient inductive bias and poor interpretability of existing graph Transformer models in small-sample brain network analysis, and the dual prior integration design has certain methodological innovation. The KD-Brain framework achieves competitive performance on multiple mental disorder diagnosis tasks, and the interpretable analysis results are in line with neuroscience consensus, which has certain clinical application value. However, the paper still has obvious deficiencies: the experimental verification is not comprehensive enough (lack of cross-center validation, subnetwork division robustness test), the analysis of LLM prior and task performance differences is insufficient, and some experimental details and symbolic descriptions are not standardized. (e.g. Some mathematical notations (e.g., q, g in the experiment table) are not clearly defined in the text, and the description of the bidirectional convolutional spatial encoder can be more detailed.) These problems affect the completeness and persuasiveness of the work to a certain extent. On the whole, the core contribution of the paper is valuable and basically meets the acceptance standard, but it needs the authors to carefully revise and improve according to the review comments. Therefore, I give a weak accept score, and the final acceptance depends on the author’s rebuttal and revision effect.
- 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.
The comments have been largely well addressed.
Review #3
- Please describe the contribution of the paper
This paper proposes KD-Brain, a prior-inspired brain network learning framework that innovatively injects disease-specific semantic priors generated by large language models into the query vectors of the attention mechanism. By regulating the interaction distribution of subnetworks through pathological consistency constraints, the framework effectively improves the diagnostic performance and neurobiological interpretability of mental disorders under small-sample conditions.
- 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.Addressing the issue of overfitting in purely data-driven models when dealing with small-sample medical datasets, the paper proposes integrating clinical pathological knowledge as semantic priors and distributional constraints into the Transformer architecture, offering a reasonable approach for combining medical domain knowledge with deep graph learning. 2.The paper explores methods for extracting textual pathological knowledge using large language models (LLMs) and converting it into numerical priors for graph learning, making a valuable attempt to bridge the gap between natural language clinical texts and deep graph neural networks. 3.The model demonstrated good generalization performance across three distinct psychiatric diagnosis tasks (ASD, BD, MDD), and the extracted key brain regions (e.g., the thalamus) and functional pathways largely align with existing neuroscience literature, indicating a degree of clinical interpretability.
- 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.When generating semantic priors, the model directly inputs text descriptions containing specific disease features, which constitutes implicit label leakage. The paper lacks rigorous ablation experiments using only disease-unrelated, purely physiological text as prior information, making it difficult to rule out the possibility that the model “cheats” by relying on textual information. 2.Although the original intention of introducing medical priors was to improve generalization, when evaluating the ABIDE dataset, the authors proactively excluded data with high heterogeneity, retaining only the single NYU site. This avoidance of multi-center validation means that the framework lacks substantial proof of robustness in real clinical settings. 3.The framework introduces several key hyperparameters that determine the model’s core behavior, such as \lambda_{sp} and \beta, yet the paper provides no sensitivity analysis whatsoever. Furthermore, the pathological consistency constraint relies heavily on the interactive prior distribution generated by a black-box large model, making it difficult for readers to assess the model’s stability and applicability boundaries under different settings. 4.The paper’s core motivation is to explore higher-order interactions among brain networks, yet the baselines it compares against not only lack the latest advancements from the past two years but also completely omit hypergraph-related methods in complex higher-order brain network modeling from recent years, severely undermining the persuasiveness of its performance claims. 5.Some of the paper’s phrasing exhibits typical templated sentence structures common in large language models, resulting in superficial descriptions of certain key technical details.
- 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?
See the weaknesses.
- 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 author’s response has somewhat alleviated my concerns to a certain extent. However, there are still some reservations regarding the comparison of the latest methods and the sensitive analysis of parameters. Therefore, I have maintained the decision of “weak accept”.
Author Feedback
We thank the reviewers (R1,R2,R4) for their positive comments and clarify the main points: ①Reliability of H_sp and P_prior(R1,2,4) The LLM generates 2 priors: disorder-specific semantic prior H_sp and subnetwork interaction prior P_prior. H_sp has been verified by clinical psychologists (co-authors) to ensure consistency with clinical practice. For P_prior, under prompt-guided CoT reasoning, the LLM performs causal inference: identifying core pathological features→mapping impaired subnetworks→inferring interactions based on theories such as the triple-network model. Clinical knowledge is then converted into normalized probability values. Moreover, results indicate that the ranking of probability values rather than absolute values is required, providing a robust inductive bias. To ensure cross-model reliability, GPT-4 and DS-R1 produce highly consistent rankings of interaction strengths (Spearman correlation>0.95), suggesting P_prior captures stable pathological patterns instead of hallucinated associations. Removing H_sp or P_prior degrades performance. Due to page limits, replacing them with random priors also result in inferior performance, further validating the effectiveness of LLM-derived priors. ②Disease-Specific Prior Leakage(R1,4) H_sp encodes the disease-specific functional identities and pathological alteration of subnetworks. We clarify that H_sp is exclusively used during training as a semantic prior to guide attention learning. During testing, H_sp is removed, and model performs inference purely based on the learned subnetwork representations and attention weights. Hence, there is no information leakage. This is further supported by the ablation results: w/o H_sp<concat(H_sp,Z(q))<KD-Brain. ③The Effectiveness of Prior Injection(R1) When a purely data-driven GAT (w/o Prior) is used to learn subnetwork interactions, performance decreases (↓5.0% on ASD), although it still outperforms several SOTAs. This suggests that, under the guidance of H_sp, the model capturing more clinically meaningful pathological patterns. ④Motivations of Positional Encoding(R1) The role of H_sp is similar to positional encoding in the Transformer by breaking its inherent permutation invariance. In the context of brain networks, we argue that spatial coordinates carry limited neurobiological information; instead, the functional identity of a subnetwork serves as its more meaningful position. By injecting H_sp into the Query, we explicitly guide the attention mechanism with the subnetwork’s pathological role, essentially acting as a Functional Semantic Positional Encoding. ⑤Definition of Heterogeneous Brain Networks(R1) As mental disorders are associated with abnormal interaction patterns among subnetworks, we use “heterogeneous” to describe the functional diversity and disorder-specific interaction patterns of subnetworks, rather than the classical graph-learning definition involving multiple node/edge types. The heterogeneity enable model to identify functional pathways with high-level semantic meanings, being consistent with findings in psychiatric literature. We will explicitly clarify this in our revision. ⑥Comparison with SOTAs(R4) KD-Brain is compared with 12 methods (e.g., CAGT IJCAI2025, EAG-RS TMI2023) covering Transformer-, GNN-, and hypergraph-based methods. KD-Brain surpasses the best baseline by 5.5% ACC. Although both hypergraph method (EAG-RS) and our method model high-order interactions, ours more effectively captures disease-relevant subnetwork associations. These methods are the most relevant works available, while newer studies were either inaccessible or lacked public code. ⑦Supplement(R2,4) KD-Brain is insensitive to λ_sp and β, maintaining stable performance across a wide range of settings. Performance fluctuations on MDD likely stem from higher pathological heterogeneity. Future work will investigate subtype partitioning, different brain atlas parcellations and cross-site validation.
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.
Based on the review comments, I think it is important the authors focus their efforts in the rebuttal in putting together a persuasive presentation of the results to justify their methodological contribution. Moreover, several reviewers are concerned whether the priors (semantic) introduce too much bias and drive the results more than the architectural/topological design. Other concerns include the procedure for subject selection, potentially excluding heterogeneity in the cohort across sites and affecting the completeness of the experimental results.
- 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 main concerns raised during the review process seem to have been addressed adequately, and multiple reviewers have increased their scores to an accept based on the quality of the responses.
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.
The authors have done a good job addressing the main critiques. This work would be of interest to MICCAI.
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 reviewers mostly agree that the proposed approach has novelty and the rebuttal addressed significant questions regarding the methodology, including LLM-guided priors, disease-specific priors, and clarification of the experimental setting for comparisons.