Abstract

Early diagnosis of attention deficit hyperactivity disorder (ADHD) in children and its underlying neurobiological mechanisms have become a focal point of research. Existing AI-based diagnostic methods show promise but struggle to fully capture dynamic correlations between brain regions, limiting their clinical effectiveness. In this study, we proposed a time&frequency-dynamic functional connectivity fusion network (T&F-DFC FusionNet) based on functional near-infrared spectroscopy (fNIRS) to assist in the objective diagnosis of children with ADHD in clinical practice. The T&F-DFC FusionNet can extract the time and frequency domain features of spatial dynamic functional connectivity across channels of fNIRS data, and improve the diagnostic results by effectively fusing multi-domain features. Meanwhile, T&F-DFC FusionNet used a leave-one-ROI-out method to study specific functional brain regions with abnormal connectivity in children with ADHD to identify clinically significant biomarkers. Through a series of experiments based on clinical data, the results show that T&F-DFC FusionNet is effective in diagnosing ADHD in children, and its performance is significantly better than that of the comparison model. In addition, Notably, our findings suggest that connectivity abnormalities in the right dorsolateral prefrontal cortex and the BA 8 may be key brain regions involved in the pathogenesis of ADHD in children. In summary, this study provides new insights and methods for clinical auxiliary diagnosis and mechanism exploration of ADHD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1732_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_T&FDFC_MICCAI2025,
        author = { Chu, Mengxiang and Ma, Yunxiang and He, Xiaowei and Li, Xiao and Ren, Jiaojiao and Zhong, Zhengyu and Yu, Jingjing and Guo, Hongbo},
        title = { { T&F-DFC FusionNet: Time&Frequency-Dynamic Functional Connectivity Fusion Network for ADHD Diagnosis in Children based on fNIRS } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {626 -- 636}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The proposed T&F-DFC FusionNet model utilizes time and frequency domain dynamic functional connectivity features from fNIRS data to enhance ADHD diagnosis in children. It identifies key brain regions like the right dorsolateral prefrontal cortex and BA 8 as significant biomarkers, achieving superior diagnostic accuracy compared to traditional methods.

  • 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. This paper is well-written, and well organized.
    2. Innovative Diagnostic Model: The paper introduces the T&F-DFC FusionNet, a novel model that enhances ADHD diagnosis in children by effectively capturing dynamic functional connectivity in both time and frequency domains using fNIRS data.

    3. High Diagnostic Accuracy: The T&F-DFC FusionNet significantly outperforms traditional machine learning methods and other deep learning models, achieving a high accuracy of 88.25% in diagnosing ADHD, demonstrating its robustness and reliability.

    4. Identification of Key Brain Regions: The model uses a leave-one-ROI-out method to identify significant biomarkers, highlighting the right dorsolateral prefrontal cortex and Brodmann area 8 as crucial for ADHD classification
  • 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) Proposed architecture: no innovative model architecture was proposed. It builds on existing techniques and attempts to efficiently combine and/or refine them (2) Data Limitations: The study relies on a relatively small sample size (47 ADHD and 47 healthy control children), which may limit the generalizability of the findings. Larger and more diverse datasets could provide more robust and widely applicable results. (3) Model explainability: need external validation (leave-one-ROI-out method ) to provide interpretability, not the model itself. (4) Why the T-DFC and F-DFC modules use different methods to calculate the functional connectivity matrix? (5) In ablation study of table 2, the proposed method should not be denoted as “baseline”

  • 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 provide sufficient information for 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.

    (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 paper is well-organized, and the proposed T&F-DFC FusionNet effectively captures dynamic functional connectivity in both time and frequency domains using fNIRS data

  • 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

    In this paper, T&F-DFC FusionNet is proposed for the task of ADHD diagnosis based on fNIRS data (verbal fluency task). It argues for the inclusion of both time and frequency information from brain dynamics and conducted ablation studies to clearly show evidence for their hypothesis. Model explainability is investigated via an occlusion-based approach.

  • 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 is written in a clear and organised manner and reasonable experiments were conducted to support their arguments.

  • 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.

    There is some room for improvement in terms of reporting of results (see details below).

  • 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 provide sufficient information for 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

    Intro

    • No major issues. It was mentioned that “ADHD patients exhibit reduced dynamic variability… making stable connectivity patterns difficult to maintain” - how does reduced variability makes it less stable?

    Method

    • The architecture is well-presented and easy to understand. However, it depends on numerous hyperparameters, e.g. number of kernels in TCN, size of kernels, number of segments / number of time frames per segment. How were these numbers arrived at?

    Experiments/Results

    • Report AUC as 0.8951 or 89.51%, not 0.8951%.
    • Ablation studies clearly demonstrate the value of the T-DFC module, F-DFC module and BiLSTM. Was there any other ablation done to investigate the value of Slices FEM ?
    • What are the model sizes (i.e. number of trainable parameters) for the various methods tested?
    • Occlusion-based model explainability approaches have often been argued to lead to OOD scenarios [1]. How does the authors think this might affect the validity of the biomarker assessment?
    • Discussion about the limitations of the proposed approach is missing. Besides methodological issues, are there other datasets that could have been used as an external dataset for further validation of the identified biomarkers?
    • Have you conducted any tests for statistical significance between the results of your proposed model and other models?
    • Are there any SOTA models that could have been compared against your model, besides just using basic SVM/RF/Transformer/ConvLSTM models)?

    Minor

    • Page 5, Section 2.3, Paragraph 2, last line: realizes -> realize

    [1] https://papers.neurips.cc/paper_files/paper/2021/file/1def1713ebf17722cbe300cfc1c88558-Paper.pdf

  • 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?

    Overall, this is a tidy piece of work that is well designed and conducted without major flaws. Although the technical novelty is not particularly outstanding (i.e. using established models like TCN and conv layers ; quite a vanilla feature fusion approach), the reported improvements looks substantial. Although there are some room for improvement as mentioned above, the current manuscript could be considered for a poster presentation in the conference.

  • Reviewer confidence

    Somewhat confident (2)

  • [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

    The authors proposed a novel T&F-DFC FusionNet capable of extract time and frequency dynamic functional connectivity from fNIRS that showed remarkable results in classifying children with ADHD diagnosis. The network is composed of three main modules that where combined time and frequency features are extracted followed by a bidirectional LSTM and two fully dense layers, where classification is made. The authors validate this novel architecture on an in-house dataset composed of 47 healthy controls (HC) and 47 ADHD children and compared to other state of the art methods from the literature, which include both traditional machine learning models like SVM and Random Forest, transformers, and CNN and LST combinations. Additionally, the T&F-DFC FusionNet adopts a leave-one-ROI-out method that allows the evaluation of the contribution of abnormal functional connectivity in different brain regions to the classification of ADHD, thus providing information regarding potential novel biomarkers for ADHD.

  • 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.

    Biomarker assessment:

    • The leave-one-ROI-out analysis reveals that the proposed model manages to capture and interpret biological patterns of brain activity by identifying key brain regions that might help characterize functional connectivity anomalies in ADHD. Dynamic time and frequency fusion feature extraction:
    • The modular approach of the T&F-DFC FusionNet provide it with the ability to extract meaningful and dynamic time and frequency features that are latter exploited by the time slice feature extraction module (FEM) to construct higher-level spatial features. Additionally, each of this module’s effectiveness was validated by ablation experiments showcasing the importance of each of the modules in the final classification ability of the T&F-DFC FusionNet.
  • 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.
    • No statistical evaluation of the comparison between models: The authors provide a comparison between the performance of the proposed network with other state of the art networks; however, they do not assess if the performance difference found are statistically significant.
    • Limited information regarding how the network is trained: There is limited information regarding how the models training is performed. We only know that the models were evaluated using a five-fold cross-validation approach, with 10% of the initial data reserved for 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 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
    • Clarity issues: In the methods section, mainly section 2.1 and 2.3 would benefit by adding some more details. In section 2.1 it could be helpful to add that the TCN is applied to each channel and then combined into the 3D tensor while providing the final tensor shape (750, 22, fc). In section 2.3, ideally, I would appreciate more detail regarding the shape of the data in the intermediate processes but I believe that the final shape of the data that exits the Slices FEM module is missing (I infer that is (10,16) due to the GlobalAveragePooling 2D). Additionally, there is no mention on the size of the two dense layers.
  • 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?

    The authors proposed a novel framework to extract and classify fNRIS data. This novel approach was validated against previous state of the art models and improved the classification ability when compared to the previous models. Additionally, this model provided some biomarker assessment which can help characterize functional connectivity abnormalities in ADHD.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 sincerely thank the reviewers for their insightful comments and constructive suggestions. Below are our point-by-point responses addressing the key concerns.

  1. Innovation and Model Design (R1, R2): (1) Previous studies on ADHD diagnosis have shown that spatial connectivity in the time and frequency domains reflect different aspects of abnormal brain coordination dynamics and are crucial for characterizing the disordered patterns of children with ADHD. Thus, we proposed T&F-DFC FusionNet method, the T-DFC used the Pearson correlation coefficient to model the spatiotemporal connectivity between fNIRS channels, while the F-DFC used the Welch correlation to capture the channels connectivity in frequency domain. Then, Slice FEM and Bi-LSTM were used to fully fuse the dynamic features in the time and frequency domains, ultimately achieving an effective diagnosis of ADHD. (2) To further support clinical applications, the leave-one-ROI-out approach aimed to analyze which ROIs or channels contribute most to ADHD classification from a biological perspective, not to study model interpretability. The results were highly consistent with known ADHD-related brain regions (rDLPFC and BA8), highlighting the clinical relevance of the extracted features. (3) Since Slice FEM has been validated in prior work, we did not repeat it and focused ablation on time, frequency, and DFC mechanism to verify the proposed network. (4) We also compared our model with two validated deep learning methods (ConvLSTMwA, Transformer-T) and two classical machine learning models to demonstrate its effectiveness.

  2. Methodological Details (R2, R3): (1)The dataset was split into 72% for training, 8% for validation, and 20% for testing. The network was trained using the Adam optimizer (learning rate = 0.0001, batch size = 8 and early stopping patience = 50), and learning rate was dynamically adjusted via ReduceLROnPlateau (patience = 10, factor = 0.5). (2) Key hyperparameters, including the number of slices, frames per slice, TCN kernel sizes and number of filters in convolutional layers, were optimized using Bayesian search based on validation performance. It should be added that TCN was applied to each channel and stacked into shape (750, 22, fc); after the Slice FEM and wavelet pooling, the features became (10,16) and were passed to Bi-LSTM. The sizes of the two dense layers were 128 and 64, respectively. (3) The number of parameters is: T&F-DFC FusionNet ~1.73M, Transformer-T ~1.88M, ConvLSTMwA ~19.8M.

  3. Experimental Results (R1, R2, R3): We have performed two-sample t-tests on all five evaluation metrics. T&F-DFC FusionNet achieved an ACC of 88.25%, significantly outperforming all baselines (p < 0.05), the SEN, F1, and AUC were also significantly higher (p < 0.05), while the SPE was better than most baselines, though slightly lower than Transformer-T. Due to space limitations, these statistical results were previously omitted, but we will include a concise summary in the revised version.

  4. Discussion and dataset limitations (R2): (1)Regarding the comment that “reducing variability leads to unstable connections”, this statement is quoted from previous literature as background motivation rather than a direct result of our study, and the reasons for this phenomenon can be found in the relevant references. (2) Data from 47 children with ADHD and 47 HC children were used in this experiment. This is a comparable dataset to existing ADHD studies based on fNIRS data in the field, which supports this experiment. To reduce overfitting and improve robustness, we used 10×5-fold cross-validation and repeated all experiments with multiple seed samples. Additionally, We are currently collecting data from a second clinical center for future external validation, which will help further evaluate the generalizability of the model and the robustness of the biomarkers.

  5. Minor Issue (R1, R2): We will correct these minor issues in the subsequent revised manuscript.




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