Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Accurately predicting post-stroke motor impairment remains a challenge due to the complexity of functional recovery and its association with neuroimaging biomarkers. This study presents a deep learning (DL) framework that integrates Magnetic Resonance Imaging (MRI)-based measures such as Diffusion Tensor Imaging (DTI) metrics—fractional anisotropy (FA), mean diffusivity (MD), radial diffusivity (RD), and axial diffusivity (AD)—along with white matter (WM) and gray matter (GM) intensities to classify upper limb motor function. Unlike previous approaches, the proposed model directly extracts whole-brain volumetric features without predefined region-of-interest constraints. Feature representation is enhanced using residual connections, attention mechanisms, and Global Average Pooling (GAP), improving classification performance while maintaining computational efficiency. The ensemble framework combines six independently trained models to optimize multi-modality integration. The results demonstrate that the WM+FA combination achieved the highest accuracy (0.97), outperforming the full ensemble model (0.96). These findings exceed the performance reported in prior studies, emphasizing the effectiveness of microstructural and structural biomarkers in motor recovery prediction. This optimized DL framework has the potential to improve post-stroke motor impairment classification, supporting early rehabilitation planning, and personalized treatment strategies.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3806_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/miccai3806/MotorImpairmentPrediction

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KarRuk_AttentionBased_MICCAI2025,
        author = { Karakis, Rukiye and Gurkahraman, Kali and Mitsis, Georgios D. and Boudrias, Marie-Hèléne},
        title = { { Attention-Based Multimodal Deep Learning Model for Post-Stroke Motor Impairment 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 = {23 -- 33}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes an attention-based multimodal deep learning framework combining DTI metrics (FA, MD, RD, AD) with white and gray matter intensity without predefined ROIs. Employing residual connections, CBAM attention, and global average pooling, there is an ensemble of six models achieved 97% accuracy for upper-limb motor impairment classification. The WM + FA ensemble outperformed full integration, underscoring microstructural biomarkers’ predictive power. This approach facilitates early, personalized rehabilitation planning by automatically extracting whole‑brain volumetric features.

  • 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 whole‑brain volumetric feature extraction without ROI constraints enhances unbiased representation of critical neuroimaging biomarkers.
    2. CBAM attention module refines feature learning by focusing on distinct salient spatial and channel‑wise information.
    3. Ensemble of six independently trained models optimizes multi‑modal integration, achieving superior accuracy over individual modalities.
  • 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. Lack of independent external validation cohort undermines confidence in model’s robustness across diverse clinical settings.
    2. PCA‑derived composite motor score may obscure individual test variances, reducing granularity in clinical outcome interpretation.
    3. Weighted ensemble lacks analysis of modality‑specific importance, hindering understanding of which biomarkers drive model predictions.
  • 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.

    (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?

    Despite outstanding accuracy and innovative multimodal fusion, absence of external validation, reliance on PCA composite scores, and limited interpretability of ensemble contributions thus significantly limit real-world clinical generalizability and reproducibility.

  • 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

    This study proposes a deep learning framework to predict post-stroke upper limb motor impairment by integrating whole-brain MRI-derived features, including DTI metrics (FA, MD, RD, AD) and tissue intensities (WM, GM), without relying on predefined regions of interest. The model leverages ensemble of six models for multi-modal integration. Results show that the WM+FA input combination achieved the highest classification accuracy.

  • 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 paper addresses a clinically relevant and timely topic—predicting post-stroke motor impairment.
    • Methodology is well-detailed, with a promising use of multimodal input data.
    • We want to commend the authors for incorporating code with their submission
    • Strong and comprehensive review of prior work in the field.
  • 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.
    • A more direct comparison with Karakis et al. (2023) would strengthen the contribution, particularly in how this approach addresses the feature fusion challenge—does ensembling simply shift the issue to a later stage?
    • While the use of multiple input types is a clear advantage, the specific contribution of each modality in the ensemble models remains unclear. Clarifying information gain from each could add insight.
    • The diffusion metrics appear highly correlated—based on the results, is the inclusion of all from the start necessary?
    • I am not sure about the rationale for treating GM and WM maps separately. Since they come from the same modality, whole intensity map might be more effective (especially as for diffusion-derived metrics you have whole brain maps).
    • Including a clinical-only baseline would provide a valuable reference point and help contextualize the model’s added value over current practice.
    • Not sure where the claim from introduction ‘Achieving superior classification accuracy over conventional ML and DL models, supporting early rehabilitation planning and personalized treatment strategies’ is supported in the paper. To might knowledge you did not directly compare to previous models on you data.
    • The introduction is thorough and informative but somewhat imbalanced relative to the results and discussion sections, which deserve more space and depth.
  • 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.

    (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?

    Due to the unbalanced structure of the paper—spending 2.5 pages on the introduction and only 2 pages on results, discussion, and conclusion (including tables and figures)—the authors don’t have sufficient space to thoroughly analyze their findings. In particular, the paper misses an in-depth discussion of the advantages and disadvantages of using different input combinations, as well as more direct comparisons with prior work.

  • 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 study proposes a deep learning (DL) model to integrate dMRI, GM, and WM measures in a non-apriori manner to predict post-stroke motor impairment prediction. The authors conclude that using WM and FA measures can classify good from poor prognosis with ~97% accuracy. Overall, the study is very innovative but the approach is unclear to reproduce.

  • 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) Designing a DL model with non-apriori measures is a significant strength. 2) The study is innovative in its approach to using ensemble learning from global pooling. 3) Designing CBAM is very innovative and is a significant strength.

  • 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) It is unclear if the model is overfitting since in the absence of no regional information, MRI measures across all voxels are being considered, and since there is no feature reduction it is difficult to convince that the model might settle in relevant features. 2) The use of augmentation though a strength is also a weakness since sharpening/data augmentation relies on uncorrelated features and in the model of stroke patients, this assumption may be violated. 3) The final observation on WM+FA is very vague. In the absence of some predictive regions, it may be that the model just happened to converge on this dataset and may not reproduce in an independent dataset. 4) The presence of FA and not MD is concerning since in the stroke-affected regions, FA will be very low. What are the author’s thoughts on this? 5) A minor comment: how was the QA done on this cohort?

  • 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.

    (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?

    This is a very innovative and significant study but the lack of details of the approach dampens the enthusiasm.

  • 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

We thank the reviewers for their valuable feedback and constructive suggestions. [R1.1]Unlike early fusion methods (Karakis et al.), our ensemble approach addresses the fusion challenge by combining modality-specific models at the decision level. This ensures balanced contributions and reduces common misclassified cases—what one modality misses, another captures. We will clarify this in the Discussion. [R1.2,R3.3]We reported individual modality performances (WM: 0.91, FA: 0.89, GM: 0.83), with WM and FA emerging as dominant contributors. Since misclassified cases show low overlap, the ensemble leverages complementary strengths to reduce errors. We will revise Table 2 and the Discussion to clarify these contributions. We will also systematically log and analyze misclassified cases from each single-modality model to quantitatively confirm their low overlap and validate the ensemble’s complementary nature. [R1.3,R2.4]While diffusion metrics appear correlated, in our ensemble what matters is which cases are misclassified by each modality (R1.2). FA and MD, though related, capture distinct properties; in our cohort, lesions often spare motor pathways, so FA reductions outside these areas may not reflect motor impairment, while MD increases in non-motor regions may mislead. WM+FA yielded the best performance, supporting their clinical relevance. We will clarify these points in the Discussion and explore combinations of DTI maps in future work. [R1.4]GM and WM maps were used to allow the model to access tissue-specific features known to play distinct roles in stroke recovery, as supported by the literature. Based on our trials, using whole intensity map did not yield comparable performance to WM and GM. Instead of explicit ROI selection, this separation allows the model to focus on detailed patterns from both cortical (GM) and connectivity (WM) structures, rather than relying on global T1 intensity maps alone. [R1.5]While our model includes demographic data (age, gender, stroke time), we did not implement a clinical-only baseline. Literature shows such models typically reach 70–80% accuracy, and our results exceed this, indicating added value. We will clarify this in the Discussion and acknowledge the benefit of including a clinical-only baseline in future analyses. [R1.6]We will revise the phrase, as no direct comparison to conventional models was performed. [R1.7]We agree and will revise the Introduction to be more concise, allowing for deeper analysis and interpretation in the Results and Discussion sections. [R2.1,R3.1]To mitigate overfitting and enhance feature focus, our model integrates CBAM attention, residual connections, L2regularization, dropout, modality-specific training. Moreover, we used multiple independent datasets, providing heterogeneity. However, we did not include an external validation cohort, and we agree this limits the generalizability of our model. We will discuss this limitation in the Discussion. [R2.2]We acknowledge this concern; however, our results benefited from augmentation by improving contrast and clarity. [R2.3]We agree that generalization is a known challenge for DL models trained on limited datasets. Nevertheless, our findings regarding the contribution of WM+FA maps are consistent with literature findings linking FA and WM integrity to motor recovery (Stinear et al.2017). [R2.5]As our study relies on imaging data, QA was ensured through our standardized DTI&MRI preprocessing pipeline (Section 2.2), such as denoising, motion correction, skull stripping, and registration. No additional manual or site-level QA was applied. [R3.2]We agree that PCA-derived composite scores may reduce individual test granularity. However, given the challenges of collecting stroke cohorts and the use of different motor assessments across our datasets, PCA was applied to harmonize these measures, prior studies (Rondina et al.). We will clarify this limitation in the Discussion.



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



back to top