List of Papers Browse by Subject Areas Author List
Abstract
While imaging-genetics holds great promise for unraveling the complex interplay between brain structure and genetic variation in neurological disorders, traditional methods are limited to simplistic linear models or to black-box techniques that lack interpretability. In this paper, we present NeuroPathX, an explainable deep learning framework that uses an early fusion strategy powered by cross-attention mechanisms to capture meaningful interactions between structural variations in the brain derived from MRI and established biological pathways derived from genetics data. To enhance interpretability and robustness, we introduce two loss functions over the attention matrix – a sparsity loss that focuses on the most salient interactions and a pathway similarity loss that enforces consistent representations across the cohort. We validate NeuroPathX on both autism spectrum disorder and Alzheimer’s disease. Our results demonstrate that NeuroPathX outperforms competing baseline approaches and reveals biologically plausible associations linked to the disorder. These findings underscore the potential of NeuroPathX to advance our understanding of complex brain disorders.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3052_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jueqiw/NeuroPathX
Link to the Dataset(s)
ACE dataset: https://nda.nih.gov/edit_collection.html?id=2021
ADNI dataset: https://adni.loni.usc.edu/data-samples/adni-data/
BibTex
@InProceedings{WanJue_Learning_MICCAI2025,
author = { Wang, Jueqi AND Jacokes, Zachary AND Van Horn, John Darrell AND Schatz, Michael C. AND Pelphrey, Kevin A. AND Venkataraman, Archana},
title = { { Learning Explainable Imaging-Genetics Associations Related to a Neurological Disorder } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15970},
month = {September},
page = {347 -- 357}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.This paper proposes NeuroPathX, an explainable AI framework that uses cross-attention to capture the intricate interplay between structural brain variation and genetic influences. 2.NeuroPathX introduces a pathway similarity loss that enforces consistent representations across the cohort to enhance interpretability and robustness.
- 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.
This paper presents an explainable AI framework intergrating sMRI and genetic data. Furthermore, genetic data is represented not by individual SNPs but by the corresponding pathways, enhancing 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.
- The text descriptions in Figure 1 and the methods section frequently mismatch, requiring clearer explanations of the paper’s methodology. For instance, the authors should specify which part of Figure 1 corresponds to the pathway-guided attention layer.
- In the comparison method section, the authors erroneously describe G-MIND and SurvPATH as baselines. Additionally, the classifier used by CCA for disease detection is not specified.
- In the learned imaging-genetic associations section, the significance of Figure 2 is poorly explained, making it difficult to interpret. 4.The paper only provides values of two hyperparameters for two loss functions. It is recommended that the authors discuss the impact of hyperparameters on results, even within limited page constraints.
- 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 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 paper’s key strength lies in transforming SNPs into pathways to improve interpretability. However, the subsequent interpretability results are not very clear. Additionally, the writing throughout the paper requires improvement, particularly in aligning figures with the corresponding text descriptions.
- 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.
I read the author’s responses carefully and almost answered the questions I asked.
Review #2
- Please describe the contribution of the paper
The paper introduces NeuroPathX, a deep learning framework that integrates MRI-derived imaging features with genetics data aggregated into biologically meaningful pathways. Main contributions: 1- It employs an early-fusion strategy using a cross-attention layer to directly capture the interactions between brain regions (ROIs) and genetic pathways. However, theoretically speaking the cross attention mechanisms are widely used and this contribution is minor.
2- Two new losses: sparsity loss (to focus on the most salient imaging–genetic associations) and pathway similarity loss (to enforce consistent pathway representations across patient and control groups) are introduced to enhance the biological interpretability of the model.
3- The method is evaluated on two clinically relevant datasets and demonstrates competitive performance compared to other methods (e.g., G-MIND, SurvPATH, and CCA).
- 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.
Evaluation on two independent datasets (ACE for ASD and ADNI for AD), and Biologically Meaningful Insights, i.e., Aggregating SNPs into pathway-level features based on established databases (KEGG) allows the model to capture biologically relevant patterns.
- 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.
Limited Dataset Size: The ACE dataset, with only 165 subjects, and even the ADNI dataset may not fully capture the heterogeneity typical of clinical populations. This raises concerns regarding the generalizability of the model.
Model Complexity and Training Stability: The architecture’s sophistication introduces additional complexity. The specialized loss functions (particularly the sparsity loss computed via random subject sampling) may lead to training instability or require careful hyperparameter tuning, which is not fully explored in the paper.
Scope of Evaluation: The experiments are limited to ASD and AD. It remains unclear how the proposed method would generalize to other neurological or psychiatric disorders, which could limit its broader applicability.
Although the authors commit to releasing code upon acceptance, more comprehensive details (e.g., sensitivity analyses and impact of hyperparameter choices) would be necessary for the reproducibility of the work.
- 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 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.
(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 main concerns relate to the limited dataset size and potential challenges in training stability due to new losses.
- 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 #3
- Please describe the contribution of the paper
This paper proposes NeuroPathX, a deep learning framework that integrates MRI-derived brain structural features with genetic pathway information using an early fusion strategy and cross-attention mechanisms. The model includes two interpretability-driven loss functions: a sparsity loss to highlight salient pathway–region interactions, and a pathway similarity loss to enforce consistency across patient cohorts. NeuroPathX is evaluated on ASD and AD datasets and shows improved classification performance and biologically plausible findings.
- 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.
Innovative cross-modal fusion: Early fusion with cross-attention effectively models interactions between brain regions and biological pathways through genetics. Pathway-level interpretation: Using pathways (rather than SNPs) improves biological relevance and interpretability. Interpretability-driven loss design: Sparsity and similarity losses enhance clarity without sacrificing classification performance. Cross-disorder robustness: Model performs well across both ASD and AD, suggesting potential robustness and generalizability. Biologically grounded findings: Pathway–region associations align with known disease biology, strengthening the credibility of the findings. Works with modest sample sizes: Practical for rare or hard-to-collect multimodal datasets. Strong evaluation: The comprehensive evaluation shows that NeuroPathX outperforms competing baseline approaches in accuracy and specificity for both datasets.
- 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.
Limited cohort size: Small datasets (165 ASD, 438 AD) may limit broader generalization. Sensitivity tradeoff: While the model achieves high accuracy and specificity, there is a notable drop in sensitivity for both cohorts, but particularly for the ADNI dataset, which the authors acknowledge but do not fully address. Static hyperparameters: Fixed settings across disorders may reduce model optimization at the trade off of generalizability. Sparse baselines: While the authors compare with some related methods, a more comprehensive comparison with state-of-the-art methods in both imaging genetics and multimodal learning would strengthen the paper. Architecture choices underexplained: The paper could benefit from more justification for the chosen architecture and encoding dimensions (e.g., why d_q = 32, d_k = 4, d_v = 8). Clinical heterogeneity unaddressed: Variation within cohorts (e.g., ASD subtypes) is not accounted for. This is especially relevant for disorders like ASD that present with considerable clinical variability. Limited biological validation: Pathway findings are literature-consistent, but lack deeper biological validation (e.g., enrichment analysis). Translation of their findings is incomplete.
- 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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.
(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 leverages early fusion and novel loss functions to uncover biologically plausible associations across disparate disorders.
- 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
Author Feedback
We thank the reviewers for their careful and constructive evaluations. All three reviewers appreciated the novelty of NeuroPathX and its ability to find cohort-level links between brain ROIs and pathway-level genetics. We respond to the major critiques by the reviewers below.
Unclear Novelty (Rev 2): The key innovation of NeuroPathX is our strategy to directly link brain ROIs and genetic pathways at the cohort level by regularizing the attention weights; such links are lacking in prior imaging-genetics works (e.g., G-MIND, SurvPATH). With regard to the cross-attention mechanism itself, most existing methods output one attention map per subject. These attention maps vary considerably and cannot answer the clinical question: “Which Brain ROI-pathway links consistently distinguish patients from controls?” We fill this gap with two new, class-level losses on the attention matrix: pathway similarity and sparsity.
Impact of the Hyperparameters (All Revs): NeuroPathX relies on two sets of hyperparameters that control (1) the network architecture (d_q, d_k, d_v, d_qk), and (2) the loss function (pathway similarity λ_path, sparsity λ_sp). A grid search on the autism (ACE) dataset was used to select the architecture values d_q=32, d_k=4, d_v=8, and d_qk=32. Notably, we determined that the model architecture must remain small to avoid overfitting on small sample sizes of N = O(100-500) but is not sensitive to the precise values. Our grid search confirmed that the loss hyperparameters also influence performance. We set them to λ_path=10^-3 and λ_sp=10^-6 to learn stable imaging-genetic associations from the attention matrix. All hyperparameters are directly transferred to ADNI without re-tuning. Our experiments on ADNI indicate that they provide a robust default setting for future studies. Code will be public on acceptance.
Small Datasets (Revs 2 & 3): While the ACE (N = 165) and ADNI (N = 438) are modest, they reflect the typical sample sizes of public datasets with both brain imaging and genome-wide SNPs. Showing that NeuroPathX yields stable, cohort-level imaging-genetic associations under these real-world constraints underscores its practical value. Extending NeuroPathX to capture ASD heterogeneity and applying it to other disorders are key directions of future work.
Details of the Comparison Methods (Revs 1 & 3): We apologize for the lack of clarity. SurvPATH (CVPR 2024) is a SOTA method for imaging-genetics. Likewise, G-MIND (2019) is another DL baseline that projects the imaging and genetics data into a shared latent space. We select CCA as our non-DL baseline due to its popularity in multimodal analyses. We believe these three baselines provide a broad sampling of current approaches in the imaging-genetics space. For CCA, we project z-scored imaging and genetic pathway features into a 10-D canonical space with linear CCA. These embeddings are concatenated and input to a logistic regression classifier. These details will be added to Section 4.
Sensitivity Tradeoff (Rev 3): While the reviewer is correct that our current operating point favours specificity over sensitivity, NeuroPathX still achieves statistically higher accuracy on both ACE and ADNI over all comparison methods. SurvPATH performs well on ADNI but is among the lowest performing on ACE, suggesting that it struggles on smaller datasets, perhaps due to its more complex architecture. Beyond quantitative performance, NeuroPathX is the only model that produces cohort-level brain ROI-pathway association. In contrast, SurvPATH provides only subject-level attributions, and G-MIND does not directly link the modalities.
Paper Clarity (Rev 1): We apologize for the mismatch between text and graphics. We will update Fig. 1 to colour-code and label the pathway-guided attention block, with a pointer to it in Section 2. Fig. 2 will be redrawn with a legend distinguishing brain ROIs from genetic pathways, making the cohort-level association map interpretable.
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”.
N/A
- 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