List of Papers Browse by Subject Areas Author List
Abstract
Neuropathic pain, affecting up to 10% of adults, remains difficult to treat due to limited therapeutic efficacy and tolerability. Although resting-state functional MRI (rs-fMRI) is a promising non-invasive measurement of brain biomarkers to predict drug response in therapeutic development, the complexity of fMRI demands machine learn- ing models with substantial capacity. However, extreme data scarcity in neuropathic pain research limits the application of high-capacity models. To address the challenge of data scarcity, we propose FMM${TC}$, a Foundation-Model-boosted Multimodal learning framework for fMRI-based neuropathic pain drug response prediction, which leverages both internal multimodal information in pain-specific data and external knowledge from large pain-agnostic data. Specifically, to maximize the value of limited pain-specific data, FMM${TC}$ integrates complementary informa- tion from two rs-fMRI modalities: Time series and functional Connectivity. FMM${TC}$ is further boosted by an fMRI foundation model with its ex- ternal knowledge from extensive pain-agnostic fMRI datasets enriching limited pain-specific information. Evaluations with an in-house dataset and a public dataset from OpenNeuro demonstrate FMM${TC}$’s superior representation ability, generalizability, and cross-dataset adaptability over existing unimodal fMRI models that only consider one of the rs-fMRI modalities. The ablation study validates the effectiveness of multimodal learning and foundation-model-powered external knowledge transfer in FMM${TC}$. An integrated gradient-based interpretation study explains how FMM${TC}$’s cross-dataset dynamic behaviors enhance its adaptability. In conclusion, FMM$_{TC}$ boosts clinical trials in neuropathic pain therapeutic development by accurately predicting drug responses to improve the participant stratification efficiency.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1399_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Shef-AIRE/FMM_TC
Link to the Dataset(s)
OpenNeuro DS000208: https://openneuro.org/datasets/ds000208/versions/1.0.1
BibTex
@InProceedings{FanWen_FoundationModelBoosted_MICCAI2025,
author = { Fan, Wenrui and Rizky, L. M. Riza and Zhang, Jiayang and Chen, Chen and Lu, Haiping and Teh, Kevin and Selvarajah, Dinesh and Zhou, Shuo},
title = { { Foundation-Model-Boosted Multimodal Learning for fMRI-based Neuropathic Pain Drug Response Prediction } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15974},
month = {September},
page = {241 -- 251}
}
Reviews
Review #1
- Please describe the contribution of the paper
FMMTC combines time‑series and functional‑connectivity streams—encoded by a frozen transformer‑based foundation model (BrainLM) and a trainable ResNet‑18 CNN, respectively—and simple feature concatenation to overcome data scarcity in neuropathic‑pain rs‑fMRI. By transferring external knowledge from large, pain‑agnostic datasets, it achieves good gains for MCC and AUROC, with strong cross‑dataset generalization on both in‑house and public cohorts.
- 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.
- Novel multimodal integration of rs‑fMRI time‑series and functional connectivity via dual‑stream encoders boosts predictive power.
- Transfer of external knowledge from massive pain‑agnostic fMRI pretraining reduces overfitting on neuropathic pain datasets.
- Rigorous cross‑dataset evaluation yields high MCC and AUROC, proving clinical stratification feasibility in neuropathic pain.
- 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.
- Fusion simplicity: simple concatenation may underutilize complex interactions between temporal and spatial connectivity modalities significantly.
- Pretrained transformer frozen layers prevent model adaptation to neuropathic pain nuances and subtle dataset-specific features.
- Binary classification simplifies complex drug response dynamics, ignoring dose–response curves, temporal patterns, continuous outcome variations.
- 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 MCC and AUROC improvements validate the foundation-model–boosted multimodal approach, demonstrating practical clinical stratification potential; the novel fusion of time-series and connectivity enhances representation, and cross-dataset robustness is notable; however, small cohorts, simple fusion, and limited interpretability requires further validation and scalability assessments.
- 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 #2
- Please describe the contribution of the paper
Adapting fMRI foundation model (BrainLM) to enhance the temporal feature extraction for neuropathic pain drug response prediction. The introduction of foundation model ease the over-fitting problem in small dataset, commonly seem in this task. The combination of time-series and functional brain network representations improves the predictability and generaliability across two datasets.
- 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.
Novel application of advanced fMRI foundation model in the neuropathic pain drug response prediction task, which is clinically significant. The performance increase is promising and significant. The evaluation is comprehensive. The introduction of foundation model ease the over-fitting problem in small dataset, commonly seem in this task. The combination of time-series and functional brain network representations improves the predictability and generaliability across two datasets. The paper is easy to follow and technically sound.
- 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.
- not mention about the code availability
- wondering why the authors did not explore graph convolutional network for functional network processing. CNN is not specifically designed for graph analysis and could not properly handle the complex topology
- Currently the functional network branch is not benefiting from the foundation model. It is possible to construct the functional networks by correlating the the temporal features from BrainLM.
- 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?
See major strengths
- 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
Introduction of a FM pipeline for processing fMRI data and performing classification to predict drug response in therapeutic development.
- 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.
Nicely written paper with a very clean structure; I really liked the idea of processing time courses and functional connectivity separately in two streams. I had never seen it before. Overall, I like the paper; a few minors and majors below.
- 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.
Major:
- What I would like to see discussed/shown here, or at least an acknowledgment of it, is performance accuracy beyond binary classification and interpretability: Would it be possible to gain some insights from the functional connectivity about which brain areas are the most important for classification?
- You stated that the FM is trained on “large pain-agnostic data”, which could be misleading unless your reference is about you use a pre-trained BrainLM, then, it may be better to refine this and be explicit.
- Fig. 1 – I can’t find easily the section where you describe the “Foundation model”, blue block on the top-left. Is it BrainLM? Add a section under “Detailed design of FMMT C”. If you describe it in the section “FM-powered external knowledge transfer”, then be explicit and move it below. Also, replace “Foundation model” with BrainLM in the figure, if this is the case.
- 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
Minor:
- Fig. 1 – Since you use concatenation for Et and Ec, I would probably simply report that in the figure, instead of the word “Feature Fusion”. I understand you tried multiple methods, but if the best is concat, write that.
- Pag.3 – change “professional” to something else, like well-known or consolidated.
- Pag.4 – “we” need uppercase.
- Would it work differently with a different atlas? What is the reason for using A424?
- Pag.5 – Which steps did you use within fMRI prep? Just add the list of pre-processing steps.
- Pag.6 – The expression “Out-of-domain” is a bit tricky. I understand you are training on a drug and test on the other, but I would naively say it’s the same domain. On this test, did you split the dataset like “train on OpenNeuro and test on the in-house”? If yes, maybe be explicit.
- 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?
I liked the idea, the structure of the experiment (and the paper overall), nicely presented.
- 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
Thank you to all reviewers for agreeing that this work is novel and interesting. We hereby reply to some of their concerns.
Q1: Alternative network design (R1, R3):
- Why CNN instead of GNN? (R1): We regard CNN as an optional FC encoder by the previous research in using AI for neuropathic pain drug response that used CNN to process FC data (doi.org/10.1007/s12021-022-09603-5). We finally adopted CNN because CNN achieved SOTA performance among all other alternative FC encoders in our preliminary study (including GNN-based models). This could be due to GNN models still suffering from overfitting with only 40-50 examples in training.
- More effective feature fusion (R1, R3): We have tried several commonly used methods in deep learning. However, due to data scarcity, the more complex the pipeline, the more severe the overfitting. Thus, we choose the simple but effective concatenation here. We have also justified the design of choice of concatenation in our paper. In addition, developing a simple, clinically inspired fusion method to model the interaction between temporal and spatial features will be one of our future research directions for this work.
- Frozen pretrained transformer layers in BrainLM (R3): We acknowledge that unfreezing some parameters in the pretrained model can help enhance the performance. However, with only 40-50 examples in fine-tuning, fully or partially unfreezing BrainLM increases the number of trainable parameters, leading to more severe overfitting. We will try more advanced techniques, like LoRA, to further improve the model’s performance in future research.
Q2: More clinical analysis and insights (R2, R3):
- Key brain regions for drug response (R2): Theoretically, through the integrated gradients (IG), we can score the importance of each dimension of the input functional connectivity matrices. However, this method requires a massive amount of GPU memory to be applied end-to-end to a deep neural network. Thus, we only interpret the final prediction layer to illustrate the model’s reliance on two modalities and its adaptability. We will extend our research with more advanced hardware to have more clinical insight into the model and experiments.
- Drug response dynamics (R3): We acknowledge that the drug response problem is much more complex than a binary classification problem. However, we need completely different types of clinical data (like protein data/omics data, etc) for an in-depth understanding of drug dynamics, which we don’t have. Instead, we focus more on accelerating the early stages of clinical trials of drugs, such as the participant recruitment stage. Thus, we don’t explore the mechanism of drugs, and we only care about whether a participant will respond to a drug.
Q3: Other concerns:
- fMRI preprocessing and atlas choice (R2): The full list of pre-processing can refer to fMRIPrep’s official document. We use the Docker file of fMRIPrep and use the default parameters and steps. In addition, the choice of an atlas will affect performance according to our experience. In this research, we choose A424 to align with the pretrained BrainLM model.
- Code availability (R1): The code will be released with the camera-ready version of this paper.
- Other minor language problems from R2 will be refined in the camera-ready paper as well.
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”.
N/A