List of Papers Browse by Subject Areas Author List
Abstract
Functional near-infrared spectroscopy (fNIRS)-based diagnosis of childhood attention-deficit/hyperactivity disorder (ADHD) is limited by spatiotemporal dynamic decoupling and neglected physiological coupling between oxyhemoglobin (HbO2) and deoxyhemoglobin (HbR). To address these issues, we propose a Spatiotemporal Heatmap Feature Fusion Network (STHFFN) for modeling abnormal neurovascular trajectories and bilateral hemoglobin interactions in ADHD. STHFFN starts with the Spatiotemporal Heatmap Construction to convert discrete fNIRS signals into anatomically continuous cortical heatmaps to preserve brain topology and provide a physiological basis for feature learning. Based on this structured representation, the Parallel Spatiotemporal attention module uses a bi-branch structure to jointly model spatial co-activation patterns and temporal hemodynamic changes, enabling unified spatiotemporal feature learning instead of separate modeling. The Bilateral Hemoglobin Complementarity explicitly captures spatiotemporal interactions between HbO2 and HbR, improving sensitivity to pathological neurovascular decoupling. The Memory Pool mechanism further aggregates long-range temporal dynamics to characterize persistent attentional instability, and multimodal features are adaptively fused for final diagnostic prediction. Extensive experiments on 47 children with ADHD and 47 healthy controls show that STHFFN achieves 92.57% ACC and 94.25% AUC, outperforming eight representative baseline models. By integrating anatomical continuity, spatiotemporal dynamics, and physiological coupling, STHFFN provides an effect approach for fNIRS-based auxiliary diagnosis of children with ADHD. Code is at https://github.com/wangbaobao7778-cyber/STHFFN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1519_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_HbO2_MICCAI2026,
author = { Chu, Mengxiang AND Wang, Xichen AND Guo, Hongbo AND Cheng, Xingxing AND Li, Xiaohan AND He, Xiaowei AND Li, Xiao AND Ren, Jiaojiao AND Yu, Jingjing},
title = { { HbO2 and HbR Spatiotemporal Heatmap Feature Fusion Network for ADHD Diagnosis in Children } },
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
This paper proposes the STHFFN for fNIRS-based diagnosis of childhood ADHD to address two key limitations: spatiotemporal dynamic decoupling and the ignorance of the physiological coupling between HbO₂ and HbR. The authors construct cortical spatiotemporal heatmaps, using parallel spatiotemporal attention and bilateral hemoglobin complementation, and aggregating long-range dynamics. Experiments on 47 ADHD and 47 healthy subjects show it outperforms baselines.
- 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 converts discrete fNIRS channels into structured cortical spatiotemporal heatmaps, which explicitly preserves the anatomical spatial topology of the brain. 2.This paper proposes a parallel spatial-temporal attention mechanism to model local brain abnormalities and dynamic temporal patterns simultaneously. 3.The model fuses both HbO and HbR signals in a complementary way, which is biologically original. By using two physiological signals instead of a single one, it captures more comprehensive brain activity patterns, leading to higher sensitivity to the weak neurovascular abnormalities in ADHD.
- 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.The core technical pipeline lacks sufficient innovation. Similar design of such a combination (spatiotemporal heatmap representation + parallel spatial-temporal attention + dual-modal hemoglobin fusion) have been explored in prior studies. 2.The study uses a small, single-site dataset without cross-center or cross-scanner validation. Generalization is critical. This paper does not test performance on external data. 3.Recent reviews have emphasized that fNIRS biomarkers for neurodevelopmental disorders require multi-center validation to avoid overfitting and ensure reliability. The current dataset size and homogeneity limit the reliability of conclusions. 4.The authors should systematically quantify fNIRS activation patterns and linked them to ADHD pathophysiology, providing quantitative neurobiological interpretation.
- 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.
- 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?
My overall score reflects a balanced assessment: this paper addresses a clinically meaningful problem with a solid, well-designed method and convincing experimental results, but it has moderate limitations in novelty, generalization, and clinical validation.
- 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.
Thank you for your revision. Your updates have partially addressed my concerns. Generalization remains a practical challenge. Without further supporting evidence, the generalization performance of this work cannot be fully verified, which also limits its practical deployment and real-world application. I suggest you further emphasize the differences, unique strengths and core contributions of this study compared with existing research.
Review #2
- Please describe the contribution of the paper
The paper proposes a DNN for anomaly detection in neurovascular data from special populations, which present with distinct physiological signals.
- 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 parameter sensitivity analysis is very helpful in demonstrating the justification behind the architectural decisions.
-
The paper is very well organized, and thus easy to follow.
-
- 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 paper acknowledges the single-center cohort drawback, but remediating follow ups are not discussed.
-
fNIRS was acquired at 10Hz, which is at the lower end of signal acquisation rate. The utility, or lack thereof, of a greater frequency is not discussed.
-
Figure 1 is constructed well, but it would be helpful to make clear how the different modules relate to each other as that is not immediately apparent.
-
- 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?
The paper presents a well-written and thoroughly validated approach.
- 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 the authors do not discuss remediation for single source dataset and generalizability concerns, e.g., leave one out CV, test time augmentation (additive Gaussian/Pink noise at a realistic SNR, random cropping, small amplitude jitter, etc.), or reporting calibration and confidence. A single source dataset is defensible because the dataset concerns the population of interest, but the generalizability concerns are not secondary because of the method’s clinical setting.
Review #3
- Please describe the contribution of the paper
The paper proposes STHFFN for fNIRS-based childhood ADHD diagnosis. The framework converts discrete 22-channel fNIRS signals into anatomically continuous spatiotemporal heatmaps via MNI coordinate mapping and cubic spline interpolation, enabling spatial convolution with anatomical meaning. A parallel spatiotemporal attention module jointly models spatial co-activation and temporal dynamics through two symmetric branches to avoid dimensional priority bias. A bilateral hemoglobin complementarity module captures the physiological coupling between HbO2 and HbR via cross-modal attention. A memory pool mechanism aggregates features across sliding windows to characterize long-range temporal dynamics. The framework is validated on 47 ADHD children and 47 healthy controls, achieving 92.57% ACC.
- 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 motivation is specific and grounded in neuroscience. The two targeted problems, spatiotemporal decoupling in existing methods and the neglected physiological coupling between HbO2 and HbR, are concrete limitations in fNIRS-based ADHD research. The HbO2/HbR coupling disruption in ADHD is supported by neurovascular coupling literature. 2.Each proposed module directly addresses a stated problem. P-ST addresses spatiotemporal decoupling, BHC addresses HbO2/HbR coupling, and MP addresses sustained attentional instability. There is no redundant component. 3.The experimental design is thorough. Five-fold cross-validation across 60 independent trials with standard deviation reporting, eight baselines, component-wise ablation, and hyperparameter sensitivity analysis provide a comprehensive evaluation within the page limit.
- 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.The memory pool (MP) shows the largest contribution in the ablation study, with ACC dropping by 5.26% when removed, compared to 2.63% for P-ST and 4.27% for BHC. However, the motivation section devotes most of its emphasis to the physiological coupling between HbO2 and HbR, with long-range temporal dynamics only briefly mentioned near the end of the introduction. The experimental results suggest that capturing sustained attentional instability over time may be more critical for ADHD classification than the spatiotemporal or coupling modeling. Adjusting the narrative to better reflect this finding would strengthen the paper. 2.The comparison between cubic spline interpolation and zero padding (92.57% vs 66.38%) shows a dramatic performance gap. It would be helpful if the authors could discuss why zero padding leads to such severe degradation. 3.The contribution of STHC is difficult to evaluate, since the input representation differs fundamentally from other models and cannot be directly swapped. To better understand its role, a simple experiment would be informative: compare a basic classifier (e.g., a linear layer or shallow MLP) applied to the STHC heatmap representation against the same classifier applied to the raw T×22 time-series input. This would reveal whether the advantage emerges at the input representation level itself, before any downstream modules are involved. Regardless of the outcome, such an analysis would be valuable.
Minor Weakness: 1.Some statements in the abstract and introduction are overly strong. For example, “robust neural decoding inherently requires explicitly modeling the physiological interdependence of local hemodynamics” and “this naive modality isolation entirely ignores the synergistic physiological coupling between HbO2 and HbR.” The paper’s own results do not fully support such absolute claims. For instance, when (P-ST+BHC) layers are set to 4, performance falls below CT-Net [16], indicating that the proposed modules do not universally guarantee improvement. Softening these statements would better align with the empirical evidence. 2.The odd-even frame separation data augmentation strategy is referenced but the paper does not explicitly state whether the five-fold cross-validation split is performed at the subject level before augmentation. Given the small sample size (47+47) and the high reported accuracy (92.57%), clarifying this detail would help confirm that no information leakage occurs between training and test sets. 3.In the figures, “Conv1d” is split across lines, reducing readability.
- 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.
- 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?
1.The paper presents a well-motivated framework with strong neuroscience grounding. The two core problems, spatiotemporal decoupling and neglected HbO2/HbR coupling, are concrete and specific to the fNIRS-ADHD domain. Each proposed module directly addresses a stated problem, and the experimental design is thorough with proper cross-validation, multiple baselines, ablation, and hyperparameter analysis. 2.The main factor preventing a higher score is the misalignment between narrative emphasis and experimental evidence. The ablation study reveals that the memory pool contributes most to performance, yet the motivation focuses primarily on HbO2/HbR coupling. This suggests the paper may not fully recognize its own most valuable finding. Additionally, the STHC representation likely plays an important foundational role, as all ablation variants outperform most baselines, but its contribution is not directly analyzed. 3.These are not fundamental flaws but rather missed opportunities to better present and understand the work’s contributions. The method is sound, the experiments are well executed, and the results are convincing. I recommend weak accept, with the suggestion that the authors adjust the narrative weight to reflect the ablation findings, clarify the data augmentation protocol, and consider simple experiments to evaluate the STHC contribution independently.
- 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.
All my concerns have been resolved.
Author Feedback
We sincerely thank all reviewers and the AC for their constructive suggestions. Below are point-by-point responses to key concerns. 1.Innovation (AC, R1) (1) While similar techniques have been explored in other fields, our core innovation of our work is an ADHD pathology-specific integrated framework, rather than standalone modules. Unlike generic spatiotemporal modeling methods, our end-to-end pipeline is designed around three core ADHD pathological features: abnormal prefrontal spatiotemporal activation, impaired neurovascular coupling, and sustained attentional instability. These pathological patterns are jointly modeled in a unified framework, forming a dedicated solution for fNIRS-based ADHD diagnosis. This pathology-driven, task-customized design fundamentally distinguishes our work from existing generic fusion models. (2) Thank you for pointing out the insufficient motivation for the MP module. We will revise the Introduction to clarify that MP is designed to capture ADHD-related sustained attentional deficits, consistent with its dominant contribution in ablation experiments. 2.Pathophysiological Link (AC, R1) Due to space limitations, we prioritized presenting the method and core performance results. We have completed quantitative analysis of brain activation across six ROIs, identifying rFPC, mFPC, and BA8 as core pathological hubs with abnormal task-state activation and neurovascular uncoupling, consistent with established ADHD literature. We will add these results and neurobiological interpretations in the revised Discussion to establish a clear link between fNIRS signals and ADHD pathophysiology. 3.External Validation and Generalizability (R1, R2) We fully appreciate the critical value of multi-center validation for generalizability. However, fNIRS-ADHD datasets are extremely scarce, and recent high-quality studies (Bian et al., 2025; Gao et al., 2025) used similar-sized single-center datasets. we use subject-level 5-fold cross-validation, and apply data augmentation only after train-test splitting to eliminate data leakage and overfitting. This rigorous design ensures the reliability of our conclusions despite the single-center setting. We will further discuss this limitation and our ongoing multi-center collaboration plans in the revision. 4.Experiments and Results (R2, R3) (1) The hemodynamic response time constant is 2–6 s (~1Hz), so 10Hz complies with the Nyquist theorem and is standard in fNIRS research. Higher sampling rates increase data volume but introduce high-frequency noise. We will add sampling rate rationale in the Discussion. (2) Cubic spline interpolation is superior to zero-padding, as it preserves brain activation anatomical continuity, maintains cortical topology based on MNI coordinates, and avoids biologically invalid spatial artifacts. We will add this analysis in Section 4.(3) Due to space limitations, we did not include fusion visualizations. Full-model performance and significant ACC degradation after BHC ablation fully verify that BHC effectively extracts the physiological coupling between HbO2 and HbR, and provides critical complementary features for classification. All fusion evidence is supported by existing experimental results, with no additional data or figures required. (4) We did not perform independent ablation on STHC, since it is the foundational input representation module of the entire framework. Removing STHC would break the pipeline and invalidate subsequent spatiotemporal modeling and fusion. Its effectiveness is verified by full-model performance and supported by recognized topology-aware learning theories in neuroimaging (Pouya et al., 2016; Meng et al., 2024). We will add relevant literature to explain its necessity. 5.Minor Issues (R2, R3) We will revise the Abstract and Introduction with more cautious, neutral wording; optimize Figure 1 to clarify module connections and data flow; and fix the “Conv1d” line-break issue for better readability.
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.
The bivariate nature of fNIRS channels means that in order to afford some statement about brain function, and in this case about functional connectivity, fNIRS researchers need to concililate the outcomes the two Hb species. Here, the proposal to achieve such conciliation is a deep ANN whose salient particulars are an attention module to capture spatial correlations and temporal dynamics that works over spatially interpolated responses (heatmaps), and a memory pool to disocciate transients from steady components in the response.
Overall a draft that is welcomed by all reviewers and myself with a lot of strengths; good experimentation (R3), good organization (R2/myself), good biological motivation (R1/R3), etc.
Among the weaknesses, I think R1 hits the nail with two specific comments; the potentially compromised novelty (this is not the only model out there to yield a fused answer from the spatiotemporal heatmaps using ANNs), and the lack of systematic link between the fNIRS observed responses and the ADHD pathophysiology. R2 and R3 also express some model specific concerns that specially have repercussions on insffuciently supported claims (R3). To this latter, I add of my own; we are promised fusioned outcomes, but unless I’m missing something very obvious we are not shown any evidence of such. The authors demands a leap of faith here.
- 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.
The post-rebuttal landscape suggests that the rebuttal did a good job including Reviewer #3 upgrading to accept. There appears to be overlap with submissions #1555 and #1557.
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 acceptance. The paper addresses a clinically meaningful ADHD diagnosis problem using fNIRS HbO2/HbR spatiotemporal feature fusion, and the reviewers generally found the method well organized, well motivated, and experimentally convincing. The rebuttal resolved or partially addressed the main reviewer concerns, and all post-rebuttal opinions are positive. Remaining concerns about single-center data, generalization, and clinical validation should be acknowledged, but they do not outweigh the paper’s strengths.
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 should be accepted as all reviewers support acceptance after rebuttal, and the remaining concerns are mainly about generalization and presentation rather than fundamental flaws in the method or evaluation.
