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

Abstract

Deep learning–based multimodal MRI integration for Amyotrophic Lateral Sclerosis (ALS) diagnosis remains largely unexplored. We present ALSAnchorNet, a biologically guided multimodal framework that integrates resting-state fMRI, diffusion tensor imaging, and structural MRI for ALS diagnosis. We introduce Frontal Anchor Attention to encode frontal lobe vulnerability by using intra-frontal functional connectivity as an anchoring query to modulate whole-brain connectivity interactions. This anchor-guided mechanism enhances disease-relevant feature amplification during cross-modal fusion. ALSAnchorNet applies modality-specific linear projections followed by anchor-based refinement to jointly model connectivity, glymphatic, and cerebral-spinal fluid (CSF) markers. Using one of the largest ALS cohort to date of 159 ALS and 107 healthy controls, ALSAnchorNet achieves 77% classification accuracy with AUC = 0.80, outperforming single-modality models and non-anchored fusion approaches. Ablation analyses demonstrate that glymphatic and CSF markers provide complementary diagnostic value beyond connectivity features. Interpretability analysis further reveals elevated contributions from the superior frontal gyri and paracentral lobule, consistent with established motor network involvement in ALS. These results establish an important first step toward biologically grounded deep learning for multimodal ALS modeling. Future work will incorporate biofluid markers and longitudinal data for more comprehensive disease modeling.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0535_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/AQ-MedAI/ALSAnchorNet

Link to the Dataset(s)

Multimodal MRI ALS dataset: https://github.com/AQ-MedAI/ALSAnchorNet

BibTex

@InProceedings{SheXio_ALSAnchorNet_MICCAI2026,
        author = { Shen, Xiongri AND Luan, Jixin AND Hu, Xingcan AND Lu, Le AND Liu, Fuchen AND Liu, Shuangwu AND Xie, Long},
        title = { { ALSAnchorNet: Biologically Informed Multimodal MRI Fusion for Amyotrophic Lateral Sclerosis Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a novel method for ALS classification using multimodal input of FC, SC, and clinical variables. The method adds frontal anchor attention grounded on the biological knowledge of frontal lobe vulnerability. They also propose to release their dataset of 159 ALS and 107 control cases. The proposed modules are novel in terms of technology but are logical ones to solve the given problem. In addition, the study will contribute to addressing the scarcity of ALS imaging 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.

    The proposed modules for ALS classification are logical ones to solve the given problem. The major module is based on biological knowledge of frontal lobe vulnerability. In addition, the study will contribute to addressing the scarcity of ALS imaging 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.
    1. They use clinical variables including PVS, etc. (dim = 93). Details are needed on the meaning and size of the variables.
    2. Why do you use a common MLP for FC and SC? Since these are two different modalities, modality-specific encoders make more sense. There should be an ablation study on this.
    3. Five-fold cross-validation might cause bias. Since the ALS/CN ratio is imbalanced, they should apply a stratified cross-validation maintaining the ratio of the two classes.
    4. Tables 1 and 2 lack p-values comparing their method with other baselines.
    5. This is a single-center study lacking external validation.
  • 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

    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?

    I am concerned with the common encoder (#2), cross-validation (#3), and statistical comparisons (#4).

  • 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.

    Clarifications are mostly satisfactory. I am a bit worried that p-values were not mentioned for the Tables since they could be over 0.05.



Review #2

  • Please describe the contribution of the paper

    This paper proposes ALSAnchorNet, a multimodal framework for classifying ALS and healthy controls. The claimed contributions are threefold. First, the paper introduces a multimodal framework for ALS versus HC classification, and it achieves SOTA performance. Second, it proposes a disease-guided attention mechanism that leverages intra-frontal connectivity to modulate whole-brain interactions. Third, the authors state that they will make their large ALS dataset publicly available upon acceptance.

  • 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 study is evaluated on one of the largest ALS cohorts reported to date. 2.The model architecture and input design are informed by prior biological knowledge, using a cross-attention mechanism to encode frontal lobe vulnerability. 3.The framework incorporates multimodal biomarkers, which help improve diagnostic performance and provide complementary information beyond connectivity features.

  • 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.Baselines appear to be collapsing to the majority class since the recall is 1 for FC + SC only and MLP fusion. 2.Weak baseline models. Classic connectome-specific baselines, such as BrainNetCNN, BrainGNN, Brain Network Transformer, etc., are not chosen. 3.No variance reporting. Only mean 5-fold CV metrics are reported, standard deviations are not reported in tables 1 and 2. 4.Figure 1 inconsistency. FC and SC dimensions are visually different, while the text mentioned they are the same dimension. 5.Poor methodological clarity.

    • The paper states that d_c = 60,516 is the vectorized upper-triangular dimension of a 246*246 matrix, which is contradictory.
    • In the Method section, the text “partition the connectivity embeddings” is misleading since the h_p and h_p,; appear to be created using different MLPs.
    • Eq 3 and Eq 4 define K and V with dimensions d_k and d_v, which is inaccurate since K and V need to be the same dimension
    • The implementation details describe a two-layer classifier (128 -> 64 -> 2), but Eq 10 only shows a single linear layer, which is contradictory 6.One of the claimed contributions, the Fusion Refinement Encoder, appears to be simply a concatenation operation followed by an MLP, which does not seem sufficiently novel.
  • 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.

  • 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?

    The paper addresses an underexplored problem and contributes one of the larger single-center ALS multimodal MRI cohorts to date, with a promised data release, which is a meaningful contribution to the field. Integrating glymphatic and CSF biomarkers with connectomes is also underexplored in the context of ALS. However, baseline comparisons are weak, the variance in the 5-fold experiment is not reported, and the method section contains several inconsistencies.

  • 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.

    The authors honestly admitted to the methodological mistakes and promised to correct them in the camera-ready version. Baseline selection and the fusion novelty remain partially unresolved, but the application contribution and the promised data release could be meaningful to the field



Review #3

  • Please describe the contribution of the paper

    The authors introduced ALSAnchorNet, a biologically informed multimodal deep learning framework that integrates connectivity, glymphatic, and CSF biomarkers for multimodal MRI-based ALS diagnosis. They proposed Frontal Anchor Attention (FAA), a disease-guided attention mechanism that leverages intra-frontal connectivity to modulate whole-brain interactions. This design improves diagnostic accuracy and interpretability while remaining parameter-efficient compared to Transformer-based fusion.

  • 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 proposed FAA explicitly encodes known ALS vulnerability, making the model more meaningful than conventional DL fusion.
    • Their design combines connectivity with glymphatic and CSF biomarkers, demonstrating clear complementary gains beyond FC or SC alone.
    • Their approach shows increasing clinical credibility since region-level and connectivity analyses align with known motor network pathology.
  • 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 methodological contribution in this work seems incremental rather than fundamentally something new.
    • The proposed “anchor attention” is a simple adaptation of standard attention with a predefined region.
    • Additionally, fixing the frontal lobe as the anchor is not learned or validated and this may bias the proposed model and limit generalization to other ALS patterns.
    • In current work the fusion is largely MLP-based with simple concatenation and refinement. No strong comparison to more advanced cross-modal methods was reported which is very important.
    • The performed evaluation is limited to a relatively small cohort, with only moderate performance (accuracy: 77%), and without any external validation. It is unclear how well the proposed model generalizes.
  • 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

    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 is well-motivated, but the overall contribution is limited by incremental novelty and a relatively simple methodological design. The proposed FAA is not sufficiently distinct from existing attention mechanisms, and key design choices are not justified. The fusion strategy is very basic, without strong comparison to more advanced multimodal approaches. Additionally, the evaluation is constrained to a small, single-center cohort with moderate performance, and no external validation was performed which leaves generalization unclear. These concerns limit the overall impact and strength of the work in this current form.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    While the rebuttal clarifies the clinical motivation and acknowledges several limitations, it does not sufficiently address my main concerns regarding the strength of the methodological contribution, limited validation, and generalization. The proposed attention mechanism remains closely related to standard attention with a fixed anatomical prior, while the multimodal fusion itself is relatively simple. External validation is also missing, and given the moderate performance and single-center evaluation, I do not feel the work is currently strong enough for acceptance.



Author Feedback

We thank all reviewers for their constructive feedback and for recognizing our contributions: the biologically grounded design (all reviewers), one of the largest single-center ALS multimodal MRI cohort to date (R2), and the clinical credibility of our findings (R4). We address the main concerns below.

1.Methodological Clarifications (Meta, R2) (1) d_c dimension (R2): We apologize for the typo. We use the full FC/SC matrix as input. We will remove the word “upper-triangular”. (2) Modality-specific encoders (R1-Q2, R2): h_p (FC) and h_{p} (SC) are produced by separate, modality-specific MLPs, not a shared encoder. We use different colors of MLP blocks in Figure 1 to indicate this, but provided the incorrect description in the Method section. We will rewrite the corresponding method section as “modality-specific embedding via two parallel MLPs.” and update equation (1) to reflect this. (3) Eq. 3–4 (R2): We fully agree K and V matrix should share the same sequence length. We incorrectly use the same symbol for both the dimension of the K matrix and the normalization scale factor d_k, We will revise the method description to clarify. (4) Eq. 10 (R2): Equation (10) will be corrected to reflect the two-layer classifier (128→64→2) used in implementation. (5) Figure 1 (R2): FC and SC are indeed the same dimension; the visual difference was an illustration artifact and will be corrected. (6) Clinical variables (R1-Q1): The 93-dim vector comprises 3 ChP, 3 gBOLD, 70 FWF, 11 ALPS, and 6 PVS features. We will add this in Section 2. 2.Statistical Rigor (Meta, R1, R2) Our 5-fold CV was stratified by class (R1-Q3); we will state this explicitly. Standard deviations across folds and paired Wilcoxon signed-rank tests against each baseline will be added to Tables 1 and 2 in the camera-ready version.

3.Baselines (R2, R4) Recall = 1.0 on weaker baselines (R2-Q1): This reflects collapse to the majority class (ALS), which actually highlights the difficulty of the task and the value of our biologically informed design — ALSAnchorNet avoids this collapse and achieves balanced sensitivity and specificity. Connectome-specific baselines (R2-Q2, R4): We have compared SOTA methods in the ALS diagnosis task. Since the original papers did not provide specific names for their methodologies, we refer to them as CNN, GNN, and Transformer in the manuscript. Furthermore, the selected baselines include established CNN/GNN/Transformer-based frameworks, which represent the core innovations of BrainNetCNN, BrainGNN, and Brain Network Transformer, respectively. In contrast, our approach selects similar methodologies that are more specifically designed for ALS. To clarify that we have compared SOTA methods, we will update the naming of the baseline methods to ‘XXX et al.’ in the revised manuscript and include broader comparisons in the journal extension.

4.Novelty and Design (R4, R2-Q6) We acknowledge that our contribution lies more in designing an appropriate method for an important and underexplored clinical application than in proposing fundamentally new machinery. To our knowledge, this is the first framework integrating connectomics with glymphatic/CSF biomarkers for ALS, and the first to encode frontal-lobe vulnerability as an architectural prior. FAA is not generic attention — its query is anatomically constrained to the frontal subnetwork, which is precisely what lets the biological prior shape whole-brain interactions. Fixing the anchor (R4) is intentional; learning it would discard the prior knowledge motivating the design.

5.Generalization (R1-Q5, R4) We acknowledge the single-center limitation. Multimodal MRI ALS datasets are extremely scarce, which motivates our public data release of the processed data (159 ALS/107 HC). Cross-site validation will be reported in future work.

We will incorporate all clarifications and corrections in the camera-ready version.



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.

    This paper presents ALSAnchorNet, which combines an attention network for intra- and inter-frontal structural/functional connectivities with glymphatic and CSF biomarker features, to identify ALS (Amyotrophic Lateral Sclerosis). The paper lacks statistical evaluation and comparison with advanced architectures, as all reviewers consistently pointed out. Although conducting more experiments is not asked during the rebuttal phase, they should be clarified in the rebuttal. The clarification questions about the methodology part need to be answered in the rebuttal.

  • 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 rebuttal sufficiently addressed reviewers’ questions regarding the clarification of the methodology and details of the experiments.



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.

    I recommend accept: post-rebuttal R1 and R2 both raised to Accept and the rebuttal addressed the primary AC’s methodology clarifications.



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.

    This paper’s main contribution is a biologically informed multimodal ALS diagnosis framework that combines connectivity features with glymphatic and CSF biomarkers, together with a clinically meaningful dataset that reviewers viewed as valuable to the field. The rebuttal did not eliminate every concern, yet it appears to have satisfactorily clarified several technical points and preserved positive reviewer support after rebuttal. On balance, the paper’s contribution to ALS imaging analysis and multimodal modeling is sufficient to justify acceptance despite some methodological limitations.



back to top