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

Abstract

Functional near-infrared spectroscopy (fNIRS) is a promising modality for brain–computer interfaces (BCIs), yet multi-class neural decoding remains limited by physiological noise, hemodynamic response function (HRF) sluggishness, and scale-dependent spatio-temporal heterogeneity. To overcome these challenges, we propose Multi-scale Dual-domain Fusion-Based Spatio-Temporal Graph Learning (MsDFSTGL), a biologically informed hierarchical framework for fNIRS decoding. MsDFSTGL first builds a noise-resilient topological prior by modeling dual-domain HbO2-HbR neurovascular coupling via a Time-Frequency Global Correlation (TFGC) module and Graph Convolutional Networks (GCN). To address HRF sluggishness and scale heterogeneity, parallel multi-scale branches utilize scale-adaptive TFGC and Multi-Stage Differential Encoding (MSDE) to disentangle transient neural triggers from slow hemodynamics. Finally, a Multi-Scale Multi-Head Cross-Attention (MSMHCA) mechanism captures inter-scale dependencies, which are adaptively aggregated by a Cross-Scale Mixture-of-Experts (CS-MoE) to minimize inter-subject variability. Extensive evaluations on three heterogeneous fNIRS datasets demonstrate superior multi-class decoding performance, achieving accuracies of 95.51% on Mental Arithmetic, 76.85% on Unilateral Finger and Foot Tapping, and 34.82% on a challenging Action Recognition Motor Imagery dataset, with strong cross-subject generalization. By explicitly disentangling non-stationary neurovascular dynamics, MsDFSTGL establishes a biologically grounded and robust framework for complex neural decoding in motor rehabilitation. Code is available at https://github.com/astro1109/MsDFSTGL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1557_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ChuMen_Multiscale_MICCAI2026,
        author = { Chu, Mengxiang AND Cheng, Xingxing AND Wang, Xichen AND Yu, Jingjing AND Li, Xiangyu AND Du, Siyi AND He, Xiaowei AND Li, Xiao AND Guo, Hongbo},
        title = { { Multi-scale Dual-domain Fusion-Based Spatio-Temporal Graph Learning for fNIRS Signal Classification } },
        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

    The paper introduces an Multi-scale Dual-domain Fusion-Based Spatio-Temporal Graph Learning (MsDFSTGL) model for fNIRS signal classification. The four modules (TFGC,MSDE,MSMHCA and CS-MoE) along the model from input to output help utilize fusion of different features. TFGC help the integration of frequency features. MSDE helps with further fusion with first-order difference of features, MSMHCA helps with fusion among different features from different time scale and CS-MoE makes the last voting.

  • 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.
    • Though building on existing techniques of GCN, MHA and MoE, The motivation and design of the four modules are solid.
    • The evaluation with LOSO cross validation on three different datasets help supports the design.
  • 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 papers talks about “noise-resilient spatial topology modeling”, “temporal disentanglement”，etc. But there are no relevant experiment results for validating the statement. For example, easy stability analysis by adding noise to the input or simply visualizing the learned graph topology/attention/voting score . . . for checking if any interpretable patterns exist.
    • The paper lacks relevant in-depth analysis on associating numerical patterns from experiment to clinical interpretations. -There are multiple locations needs further clarification/explanation in the model design:
    • In the TFGC, how to calculate the frequency-domain Pearson correlation since they are complex vectors?
    • The signal is not stationary, it has some theoretical pitfalls to apply directly FFT on the segments (for example, one of them is the well known spectral leakage problem). In my opinion, it makes more sense to use methods such as STFT and use Magnitude-Squared Coherence (MSC) instead of pearson correlation.
    • In eq (1), if W is not symmetric, the result may not be symmetric. How do authors enforce the symmetry of Am?
    • What is the hard threshold used to control the density of Am?
    • In MSDE, what is the length used for patching？
    • In MSMHCA module, which dimension are those query tensors concatenated along?
    • It is not clear, for example, how 8 subject and 174 trials in MA ends up with input tensor of 52×2×140.Same issue for the rest two datasets.
    • For the issues in evaluation:
    • Details like stopping criteria are not mentioned.
    • Lacking relevant test for statistical significance of improvements.
    • The most recent model included in benchmark are from 2023 which is already 3 years ago. More recent methods should be included for better comparison.
  • 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
    • It may be better to include the model scale in the benchmark table (for example, number of trainable parameters).
    • Fig 2 is not very informative. Information like “model from which fold is used?” “Among the points, which are from train and which are from test?” “which class does each color associated with?”, etc. are not known.
  • 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 method includes many non-trivial tweaks and calibrations for the prediction task based on existing tools such as GCN, MHA, MoE.
    • The evaluation can be strengthened according to the above comments on the weakness.
    • The experiment protocol appears to be sound with two different types and on three datasets.
  • 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.

    Most key questions of mine have been addressed in the rebuttal. I would still recommend authors to include more recent SOTA methods in the benchmark. Also, the performance of capsule net is unreasonably low in the table, it is better to double check or at least have a discussion in the version to present.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a novel framework named MsDFSTGL for multi-class fNIRS signal classification to address the challenges of physiological noise, hemodynamic sluggishness, and inter-subject variability. The framework integrates a Time-Frequency Global Correlation module to build noise-resilient spatial topologies, employs Multi-Stage Differential Encoding to extract transient neural gradients from slow hemodynamics, and utilizes a Cross-Scale Mixture-of-Experts mechanism to adaptively aggregate scale-specific features for robust cross-subject generalization.

  • 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 presents a highly comprehensive and biologically informed architectural design that systematically addresses the fundamental limitations of fNIRS signals through specifically tailored modules like TFGC for spatial noise and MSDE for temporal sluggishness. The experimental validation utilizes three heterogeneous datasets encompassing both cognitive and complex motor tasks which demonstrates the versatility of the proposed method. Furthermore the ablation studies and leave-one-subject-out cross-validation strongly support the effectiveness of the individual components and the model’s robustness against inter-subject variability.

  • 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.

    While the framework achieves competitive results on some datasets, the absolute classification accuracy on the AR-MI dataset is only 34.82% which casts significant doubt on the model’s practical capability for complex motor imagery tasks. The framework introduces multiple parallel complex modules including TFGC, MSDE, MSMHCA, and CS-MoE, but completely omits an analysis of computational complexity, inference latency, and parameter count. For BCI applications where real-time decoding is critical, this omission is a fundamental flaw. Additionally the thresholding mechanism used to enforce sparsity in the TFGC adjacency matrix relies on empirical tuning and its robustness across different signal-to-noise ratio environments is not adequately discussed.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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 proposes a heavily cascaded architecture to address fNIRS decoding challenges, but the severe lack of computational complexity analysis makes its practical utility for real-time BCI highly questionable. Furthermore the exceptionally low accuracy on the challenging AR-MI dataset contradicts the claims of robust complex neural decoding. Therefore my initial recommendation is a Weak Reject and I expect the authors to provide a rigorous computational cost analysis and critically clarify the clinical viability of the AR-MI performance in their rebuttal.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper proposes a graph-based fNIRS signal classifier which generalizes well across datasets with different tasks and classes. It also demonstrates high inter-subject variability.

  • 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 study design is thoughtful because it uses task diverse datasets with a different number of classes, and also evaluates the method and comparative approaches with LOSO. The t-SNE visualization is also very helpful in giving a quick way to understand the methodological approach.
    • The paper compares the proposed method against general and domain-specific models, and also performs a thorough ablation study.
    • The methodology subsections are easy to follow and motivate the approach well, but could use more references.
  • 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 title is unnecessarily heavy because it emphasizes both “dual domain” and “spatio temporal” to describe the work of the same module (TFCN). Maybe it may make sense to stick with one or the other.
    • The method seems compelling, but it’s unclear how computationally expensive it is relative to other methods.
    • Minor point: please catch typos. For example, Time-Frequency Correlation in abstract, and first line on page 3.
  • 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.

    (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?

    While the approach is described with some unnecessary flourish in some places, it was investigated and presented well.

  • 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.

    This paper is now a weak accept for me because despite at least two reviewers bringing up concerns about computational complexity (R2 and myself, as far as I can tell), the authors provide numbers without additional details such as variance, bottlenecks, scaling, or device/software specifics to really ascertain the feasibility of the method for a BCI context. Wall clock time isn’t complexity analysis, and can mislead about bottlenecks.



Author Feedback

We sincerely thank all reviewers and the AC for constructive suggestions. Below are our point-by-point responses to the key concerns. 1.Physiological Plausibility (AC, R1) Our design is grounded in neurovascular coupling, temporal phase synchronization, and frequency-energy similarity of functional connectivity. TFGC constructs noise-resilient spatial topology via dual-domain HbO2/HbR modeling, alleviating connectivity degradation caused by hemodynamic delays. GCN aggregates functional connectivity, while MSDE decouples transient neural activity from slow hemodynamic responses. SOTA performance on noisy real-world fNIRS datasets and consistent cross-dataset gains demonstrate that the learned topology and attention patterns capture meaningful brain activity. 2.Model Design Details (AC, R1, R2, R3) (1) TFGC computes frequency-domain Pearson correlations using complex coefficient magnitudes to generate real-valued symmetric matrices. We appreciate the STFT+MSC suggestion and will compare it in future work.(2) Both correlation matrices are symmetric, so Am remains inherently symmetric after sigmoid activation.(3) The 10% sparsity threshold of Am has been specified in Section 3.(4) The length of patching in MSDE is {2,16}.(5) In MSMHCA, each scale generates Q, while features from other scales form K–V pairs, producing multi-scale multiple outputs. Figure 1 will be clarified.(6) Early stopping uses a patience of 10 epochs. 3.Computational Complexity (AC, R2, R3) MsDFSTGL contains 1.5M/2.9M/2.4M parameters, requires 0.37G/0.14G/0.25G MACs, and achieves 35.2/33.2/36.2 ms latency on three datasets. fNIRSNet achieves 10–12 ms, while fNIRS-T requires 45–46 ms. All tests used identical hardware. With latency below 40 ms, our model satisfies the 100 ms real-time BCI requirement. 4.Experiments & Statistical Analysis (AC, R1) (1) Continuous recordings were segmented into task trials, and each trial was treated as an independent sample. Thus, MA contains 174 trials from 8 subjects, each represented as 52 × 2 × 140 (channels × hemoglobin × time points). The same construction was applied to UFFT (20 × 2 × 256) and AR-MI (36 × 2 × 192).(2) Baselines were selected from representative 2022–2024 studies. Other 2024–2026 methods neither reported results on our datasets nor released reproducible code, preventing fair comparison.(3) Two-sample t-tests with FDR correction were conducted on all five metrics. MsDFSTGL significantly outperformed all baselines on UFFT (p<0.05), and all except fNIRSNet on MA and AR-MI.(4) For noise-resilient topology, removing TFGC reduces ACC by 6.37%, 4.45%, and 6.18%, while removing either time or frequency branches consistently degrades performance, confirming a stable dual-domain spatial prior. Removing MSDE causes ACC drops, validating effective temporal disentanglement. These results will be added in the revision 5.Clinical Value & Applications (AC, R1, R2) (1) Performance differences across datasets align with established neuroscience, indicating that the model captures genuine neural activity patterns. LOSO performance (94.23% on MA, 77.24% on UFFT) further validates cross-subject generalizability.(2) Our survey of 2012–2026 fNIRS MI studies shows limited multi-class research (40% for 4-class tasks). Our model achieves 34.82% on 6-class AR-MI (41.55% LOSO), above the 16.67% random baseline, demonstrating SOTA performance and scalability. Stable performance across 20-channel 3-class, 36-channel 6-class, and 52-channel 2-class tasks further demonstrates strong scalability. 6.Manuscript Optimization (R1, R3) (1) Figure 2 uses test samples from the fold closest to the 5-fold cross-validation mean. Fold indices, train/test splits, and class legends will be added.(2) “Dual-domain” refers to TFGC-based time–frequency modeling, while “spatio-temporal” refers to spatial GCN and temporal MSDE. These two dimensions will be clarified in the manuscript.(3) All typos and terminology inconsistencies will be corrected.



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 presents a BCI oriented classifier over fNIRS. The proposed model fuses the two Hb observations claiming to model the “neurovascular coupling” via a Time-Frequency Global Correlation (TFGC) module and Graph Convolutional Networks (GCN). While the computational latter seems correct, the physiological modelling claim is not shown but as usual in fNIRS assumed. On top of this initial model the actual classifier uses attentional mechanisms and aggregation to reach a final multi-class classification over several tasks.

    The reviewers found strengths in the solid motivation (R1), and the evaluation over a diversity of tasks (R3). The LOSO evaluation was found adequate by all three reviewers, and the abalation study was also found useful (R2/R3/myself).

    However, there were concerns with regards to the absence of statistical inference (R1/myself), the absence of a computational complexity analysis that establishes the potential latency that could be critical for BCI (R2/R3), the lack of in depth analysis (R1/myself), the emprical tuning (R2) and the presence of unsupported claims (R1/myself). In addition to this, I would also question the scalability of the approach.

    Reproducibility is compromised by the anonymization of the code.

    I’m suggesting requesting a rebuttal. Specifically, the rebutal must address:

    • Carry out the adequate statistical analysis
    • Carry out a computational complexity analysis and link it to the adequacy of the solution for BCI
    • Provide evidence of scalabilty, preferably theoretical.
    • Justify model tuning.
    • Attempt a more in depth analysis.
    • Provide access to code.
  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    In my meta-review I summarised the concerns into issues regarding statistical inference, computational complexity analysis, scalability, unjustified model tuning, and insufficient in-depth analysis. Although, the rebuttal led to Reviewer #1 upgrading to Accept, Reviewer #2 maintained his score and the original concern about scalability and statistical rigor was not fully resolved. There appears to be overlap with submissions #1519 and #1555.



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.

    Most reviewers agreed that the authors’ rebuttal addressed the key concerns raised during the review. The remaining issues, such as adding more recent baselines, checking the capsule network result, and providing a clearer discussion of computational complexity, can be addressed in the camera-ready version.



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.

    I suggest accept. All reviewers support acceptance after rebuttal, the method is well evaluated across multiple fNIRS datasets with LOSO validation, and the remaining concerns about computational complexity and benchmarking are important but not sufficient to outweigh the contribution.



back to top