List of Papers Browse by Subject Areas Author List
Abstract
Adolescent substance use initiation (SUI) risk arises from evolving interactions among brain structure, function, and environment, yet most predictive models treat connectivity as static and overlook developmental dynamics. We proposed BrainKODE (Koopman Operator for Developmental Coupling Evolution), a geometry-aware deep framework that models longitudinal structure–function coupling as a controlled latent dynamical system. Structural and functional connectomes were aligned within a shared Riemannian space to preserve developmental consistency, while adaptive cross-modal interactions captured network-specific structure–function coupling over time. A temporally constrained Koopman formulation modeled the evolution of latent representations, with contextual factors directly modulating developmental trajectories. On the ABCD cohort, BrainKODE improved future SUI prediction over competitive baselines while yielding interpretable maps of time-varying structure–function coupling and contextual influence, highlighting controlled geometric dynamics as a principled approach for forecasting adolescent SUI vulnerability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5796_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{MazBad_BrainKODE_MICCAI2026,
author = { Mazumder, Badhan AND Wiafe, Sir-Lord AND Wu, Lei AND Calhoun, Vince D. AND Ye, Dong Hye},
title = { { BrainKODE: Controlled Koopman Dynamics over Longitudinal Structure–Function Coupling for Early Prediction of Adolescent Substance Use } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a structure-function coupling deep learning framework for predicting adolescent substance use initiation (SUI) from longitudinal multimodal connectomes, combining Riemannian alignment, adaptive multiplex GNN coupling, and semigroup-constrained Koopman dynamics with contextual control inputs.
- 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.There are three core innovations in this manuscript which are all coherently motivated and cleanly composed. Riemannian longitudinal alignment, adaptive cross-modal coupling gates, and semigroup-constrained Koopman evolution. Additionally, to enforce SPD structure on SC via diffusion operators so both modalities share the same Riemannian treatment is particularly elegant. 2.The +10.4 pp sensitivity gain over the closest Koopman baseline (NeuroKoop) and the notably balanced sensitivity/specificity (85.18/85.78%) are clinically meaningful results for an imbalanced dataset (∼23% positive rate). Most competing methods sacrifice sensitivity badly under this imbalance. 3.Table 2 systematically decomposes contributions across geometry, coupling, dynamics, and context, with four well-defined research questions. The finding that removing Riemannian alignment drops accuracy by ~9 pp is compelling and convincingly justifies the geometric machinery.
- 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 cohort-level Fréchet mean G is computed across “all training subjects and visits. “ It is critical to clarify whether this anchor is recomputed strictly within each cross-validation fold’s training split, or computed once across the full dataset. 2.The dataset is substantially imbalanced (579 SUI vs. 1,908 non-SUI, roughly 1:3.3). The paper only reports accuracy and sensitivity/specificity alone; AUROC and balanced accuracy should also be reported here. 3.The experiments are conducted on a relatively small parcellation of 53 networks, yet no analysis of computational cost, memory usage, or scalability to finer parcellations (e. g. , N=200+) is provided. This is a notable practical gap, as many fMRI related applications require higher-resolution connectomes. 3.(Minor) Reference 1 and 2 are duplicated.
- 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.
(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?
1.The three core innovations — Riemannian longitudinal alignment, adaptive cross-modal coupling gates, and semigroup-constrained Koopman evolution — are coherently motivated and cleanly composed. And later ablation study removing Riemannian alignment drops accuracy by ~9 pp is compelling and convincingly justifies the geometric machinery.
- 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.
The authors has replied well to my concerns and all other reviewers’.
Review #2
- Please describe the contribution of the paper
This paper models evolving structure–function brain coupling across multiple visits in a shared Riemannian space. It incorporates contextual variables as control signals in the latent dynamics, aiming to improve both long-horizon prediction accuracy and interpretability of developmental risk trajectories.
- 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’s major strength is its well-integrated formulation of longitudinal brain risk prediction. Rather than treating structural and functional connectivity as static or independently fused inputs, it models their evolving coupling across development, which is better aligned with the neurodevelopmental nature of adolescent substance use risk.
A second strength is the geometry-aware design: mapping connectomes into a shared Riemannian space is an interesting way to improve cross-visit consistency.
The semigroup-constrained Koopman formulation is also appealing because it imposes temporally coherent multi-step evolution from baseline.
- 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’s mechanistic claims are stronger than its evidence. Although BrainKODE is framed as learning developmental structure–function dynamics, the experiments mainly show improved prediction, not that the learned coupling gates truly reflect neurobiological coupling mechanisms. The contextual module is also only weakly validated, since gradient-based importance does not demonstrate how context changes latent trajectories. In addition, the evaluation relies on 5-fold cross-validation only, without more stringent tests of site, scanner, or demographic robustness that would strengthen the clinical claim.
- 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.
(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 overall idea is coherent and technically promising, but the evidence does not fully support its stronger claims about mechanistic longitudinal structure–function dynamics. The gains are encouraging, yet the validation remains prediction-oriented, with limited proof of biological interpretability and insufficient robustness testing for a high-impact venue.
If the authors add stronger validation of the learned coupling dynamics, such as analyses showing that the gating patterns correspond to meaningful developmental changes rather than merely predictive signals. It would also benefit from more rigorous robustness experiments, for example across sites, scanners, or demographic subgroups, and from a clearer discussion of the assumptions behind the Koopman-based controlled dynamics.
- 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 have solved some major questions, so my final opinion of the paper is accept.
Review #3
- Please describe the contribution of the paper
This paper proposes BrainKODE, a model to predict adolescent substance use by modeling how brain structure and function change over time. The method combines SC and FC, models their interaction, and uses a dynamical system (Koopman operator) to predict future brain states. Context information is also included.
- 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 paper addresses an important problem: modeling longitudinal brain development instead of using static connectivity. This is well motivated. 2.The experiments include comparisons and ablation, and the performance looks better than baselines.
- 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 method assumes that brain development can be modeled as a linear dynamical system in latent space. This is a strong assumption, but the paper does not provide enough discussion or validation to support it. 2.The results are reported as mean ± std, but no statistical testing is provided to support the performance gains. 3.Evaluation is limited to one dataset. All experiments are conducted on the ABCD dataset. It is unclear whether the method generalizes to other datasets or populations.
- 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.
- 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 studies an important problem, and the longitudinal modeling idea is meaningful. The experiments also show better performance than baselines. However, there are some key concerns. The method assumes a linear dynamical system, but this is not well justified. The results do not include statistical testing, so the improvements are not fully convincing. Also, the evaluation is only on one dataset, so generalization is unclear.
- 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.
Although the rebuttal addressed several concerns, I still have some reservations. (1) The evaluation is still limited to the single ABCD cohort, so the external generalization remains unclear. (2) The current analyses do not sufficiently demonstrate that the learned coupling gates correspond to meaningful neurobiological mechanisms, and the interpretability claims are still relatively weak.
Author Feedback
We thank the AC and reviewers for their feedback. We first address the AC’s central concern, followed by reviewer-specific responses.
Validation beyond predictive accuracy [AC, R2, R4]: BrainKODE validates learned developmental dynamics through three complementary mechanisms beyond outcome prediction. (i) Trajectory reconstruction. The controlled evolution h_hat_{t+Delta} = U(Delta) h_hat_t + B c_tilde_t (Eq. 2), initialized at h_0, is constrained by L_dyn (Eq. 3) to reconstruct observed intermediate embeddings h_2 and h_4, providing label-independent validation of longitudinal trajectories. (ii) Structural constraints. U(Delta)=exp(Delta Omega) enforces the semigroup property U(Delta_1+Delta_2)=U(Delta_1)U(Delta_2), ensuring temporally consistent evolution under a shared operator. Removing Koopman dynamics or the semigroup constraint degrades performance (Tab. 2; 77.42% and 82.50% vs. 85.64%), indicating reliance on consistent dynamics rather than static representations. (iii) Biological coherence. Learned coupling patterns (Fig. 2) show systematic longitudinal differences, including strengthened CON–SMN/VSN interactions and reduced DMN integration in SUI, consistent with prior neurodevelopmental findings.
Fréchet mean / metrics / scalability / refs [R1]: The Fréchet anchor G is recomputed within each cross-validation training fold only, preventing test-data leakage. Despite class imbalance, BrainKODE achieves balanced sensitivity/specificity (85.18%/85.78%) with 85.64% balanced accuracy, exceeding all baselines in Tab.1.Computationally, the framework remains tractable through sparse multiplex graphs (k=7), a single GAT layer, and Koopman evolution on pooled visit embeddings rather than full graph tensors. Ref. [1] and [2] will be merged.
Mechanistic vs predictive [R2.1]: BrainKODE’s mechanistic claims rest on L_dyn enforced trajectory reconstruction and systematic time-evolving coupling, not on causal attribution. The γ-ablation reduces accuracy from 85.64% to 81.37% (Tab. 2), indicating that adaptive coupling is functionally necessary. Coupling patterns also evolve systematically across visits (Fig. 2), supporting that γ captures organized developmental structure. Context enters dynamics through B c_tilde_t (Eq. 2), directly modulating latent evolution within the controlled dynamical formulation.
Robustness and generalization [R2.2 / R4.3]: ABCD is a large-scale multi-site cohort spanning 21 sites and multiple scanner platforms; subject-level stratified cross-validation evaluates generalization across this heterogeneity. Moreover, we performed leave-one-site-out evaluation across multiple major ABCD sites (holding out each site entirely for testing and averaging performance across held-out sites), achieving 83.22% ± 0.11 accuracy, 82.73% ± 0.24 sensitivity, and 83.36% ± 0.10 specificity, supporting robust cross-site generalization.
Linearity assumption [R4.1]: Koopman theory states that nonlinear systems admit linear representations in sufficiently rich observable spaces. In BrainKODE, the geometry-aware multiplex encoder defines this latent space, while U(Delta) governs the resulting latent evolution. Thus, linearity enables stable and temporally consistent dynamics rather than imposing a restrictive assumption. Removing Koopman dynamics reduces accuracy to 77.42% (Tab. 2).
Statistical testing [R4.2]: Paired t-tests across the 5 CV folds against the strongest baseline NeuroKoop yielded significant improvements with p=0.016 (accuracy), p=0.006 (sensitivity), and p=0.040 (specificity). One-sided Wilcoxon signed-rank tests further yielded p=0.0312 for accuracy and sensitivity, as BrainKODE outperformed NeuroKoop on every fold; specificity showed consistent directional improvement (p=0.063; 85.78% vs. 84.91%). Tab. 2 further shows systematic degradation across ablations, confirming that gains arise from structured modeling.
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.
As acknowledged by the reviewers, the proposed framework is very novel. The one concern that I’d hope the authors to respond to is whether there are ways to validate the value of the method beyond predictive accuracy, given the focus on learning multi-modal developmental dynamics.
- 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 manuscript receives mixed rating post rebuttal. After careful consideration, my assessment is that the weaknesses outweigh the strengths. The proposal is novel, but arguably extremely complex. To this end, the experiments failed to justify the need for the over-engineering over existing SC-FC methods. In particular, only one of the baselines is a SC-FC coupling method (others are generic GNN methods) despite that there are many papers on SC-FC coupling in past MICCAIs alone, e.g., DFSC,HKC, joint GCN. Moreover, the rebuttal did not successfully address the interpretability problem, which is a critical concern for this overly complex method.
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 think that the idea is interesting. Although some weaknesses still remain, I thought that authors have provided a good discussion on this early-stage but interesting research, and have provided some convincing responses to specific critiques. I strongly encourage authors to include better clarifications and revisions as suggested by reviewers to strengthen the final version. I would argue for accept.
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.
The reviews are inclined towards acceptance. R4 still questions generality of the method, but given limited nature with datasets in medical imaging and that CV has been used for evaluation, I would recommend accepting the paper.