List of Papers Browse by Subject Areas Author List
Abstract
Understanding how individual infant brain function evolves over development requires models that capture personalized longitudinal changes in functional dynamics, rather than static connectivity estimates. However, existing data-driven approaches largely focus on longitudinal functional connectivity (FC) prediction, leaving a significant gap in mechanistic interpretability, causal analysis, and grounded support for targeted intervention in atypical brain development. To address this challenge, we propose the first Neuro-Developing Engine (NDE), a novel surrogate longitudinal model of infant brain dynamics utilizing resting-state functional MRI (rs-fMRI) data. By learning age-dependent transitions between individualized surrogate brains at the individual level, NDE not only predicts the evolution of surrogate brain dynamics across continuous ages but also identifies brain regions exhibiting strong modulatory effects during development. Specifically, 1) leveraging individualized Neural Perturbational Inference (NPI) surrogate models that capture subject-specific brain dynamics, we design a developmental residual model equipped with FiLM-conditioned generators to distill explicit age-driven modulation between younger and older brain dynamics, conditioned on graph-embedded individual identity and ages. 2) To enhance causal and age-sensitive representations, we combine causal distillation and Jacobian spectrum objectives with a soft-label contrastive loss that enforces alignment between predicted residual dynamics and age gaps. 3) To enable inverse-control analysis, we introduce region-level “counterfactual developmental knobs” that modulate residual dynamics, allowing the quantification of how age-dependent perturbations reshape FC over development. Extensive experiments demonstrate NDE’s ability to model individualized developmental dynamics, predict connectivity changes, and support counterfactual inverse-control simulation, highlighting its potential for mechanistic understanding and guiding early intervention in typical and atypical neurodevelopment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1965_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{HuDan_The_MICCAI2026,
author = { Hu, Dan AND Cheng, Jiale AND Han, Kangfu AND Wu, Zhengwang AND Wang, Li AND Lin, Weili AND Li, Gang},
title = { { The Neuro-Developing Engine: a Surrogate Longitudinal Model for Infant Brain Dynamics Via Causal Distillation } },
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 presents a highly innovative and methodologically sophisticated framework, the Neuro-Developing Engine (NDE), for modeling longitudinal brain dynamics in infancy. The core idea of learning age-conditioned transitions between individualized neural surrogates represents a significant conceptual leap beyond static or predictive connectivity modeling. The work is exceptionally well-structured, with clear motivations, a comprehensive methodological presentation, and rigorous experimental validation on two independent cohorts. The introduction of “counterfactual developmental knobs” for in-silico interventional analysis is a particularly compelling contribution that enhances the mechanistic interpretability of the model. The results are convincing, demonstrating superior performance against ablated variants and the model’s utility in identifying divergent trajectories in an at-risk population.
- 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 primary innovation is the formulation of neurodevelopment as a process of transitioning between subject-specific dynamic models, rather than predicting static functional connectivity (FC) or learning a population-average trajectory. This approach inherently captures individual differences and the evolution of dynamics, not just correlations.
- The combination of FiLM-conditioned residual dynamics, age-gap-aware gating, and causal distillation (behavior + EC + Jacobian spectrum) is well-designed and empirically validated.
- The PDE cohort demonstrates real-world utility—identifying divergent connections and proposing counterfactual interventions.
- 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 compares NDE with RNN/CNN variants of the surrogate architecture and ablations, but does not compare with other longitudinal FC prediction methods.
- Section 2.2 describes counterfactual generation from random noise, but the free-rollout FC used in evaluations (Table 1) is not fully explained. How many steps? Is the model unrolled autoregressively from an initial noise vector? The authors should clarify the procedure for generating FC_free, perhaps in a dedicated paragraph or supplementary note.
- Table 1 reports means and standard deviations, but no p-values or statistical tests are provided to confirm that NDE’s improvements over baselines (e. g. , NDE w/o C, NDE-FC) are significant.
- The “developmental knobs” are described as modulating residual dynamics (Eq. 9). However, the paper does not explain how the intervention vector g is optimized or selected for the results in Fig. 3c-e. Is g learned via gradient descent to minimize divergence? Or is it a fixed perturbation? The methodology for identifying “top FC drivers” should be detailed.
- The model trained on BCP is fine-tuned on 50% of TD infants from the PDE dataset (Fig. 3a caption says “fine-tuned on 50% TD infants”). However, the PDE dataset has only 35 TD controls. Fine-tuning on 50% of those (≈17-18 subjects) may lead to overfitting. The authors should report the number of TD subjects used and consider cross-validation or a more robust fine-tuning strategy.
- The conclusion is brief and does not discuss limitations.
- 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?
This is a high-quality paper that introduces a novel and impactful framework for modeling longitudinal brain development.
- 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
The main contribution of the paper is that their proposed method, NDE, unlike existing methods, models continuous developmental changes rather static connectivity.
- 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 major strength of the paper is the fact that the proposed method allows longitudinal modeling, unlike cross-sectional approaches of the past. Furthermore, the way that the method was developed allows it to simulate transitions between ages, rather than disconnected timepoints. The model does so while preserving subject-level individual properties.
- 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 model makes assumptions about certain neurodevelopmental patterns for the sake of stability and implementation of model, such as the fact that the brain development takes on a monotonic, plateauing curve pattern. While this may be accurate for healthy infant aging, if the model were to be adopted for other patterns such as aging across human life span, such as events that may lead to regression or nonlinear shifts, such assumption may become a limitation of the model.
- 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?
I believe the conceptual novelty of proposing a longitudinal, continuous model serves as a strong selling point of the work. Additional external validation and biological/clinical findings that further validate the predictions of the model would make the model more interpretable. Finally, some of the assumptions of the model may make it more stable and good at reproducing expected patterns, but may limit what it can accurately model.
- Reviewer confidence
Not confident (1)
- [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 paper introduces a NDE framework that models how infant brain dynamics evolve over time using resting state fMRI data where the connectivity evolves dynamically as opposed to treating it as a static measurement. It builds subject-specific surrogate brain models and learns age-driven transitions between them, while also allowingcausal analysis and counterfactual simulations on brain regions.
- 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 formulation is technically sound, especially the FiLM-conditioned residual transition module that models how brain dynamics evolve with age while preserving subject identity. This is a nice way to separate subject-specific structure from developmental effects.
2.The paper goes beyond prediction and explicitly targets causal interpretability, using Jacobian-based effective connectivity and causal distillation. This makes the model more aligned with neuroscience questions.
3.The counterfactual developmental knobs is a strong idea, as it allows simulation of how perturbing specific regions affects future connectivity.
- 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 whole framework depends on training separate surrogate models per (subject, age) and then learn transitions between them, but from a computational perspective, it seems that it is data inefficient and could scale poorly. There is no discussion of how stable these surrogates are with limited timepoints.
2.The comparisons are weak with respect to other competing methods, since Table 1 only compares against standard models like (baseline surrogate, RNN, CNN, ablations) rather than established longitudinal models like [1]. Given that [1] is already discussed in the introduction, it would have been more convincing to include it in the evaluation.
3.The FiLM-based residual transition in the paper assumes that the developmental changes are additive corrections from the source surrogate (Eq. 4–5), which is a fairly strong assumption? In early brain development where dynamics could change nonlinearly, this residual formulation might be too restrictive but this is not discussed.
[1] A few-shot learning graph multi-trajectory evolution network for forecasting multimodal baby connectivity development from a baseline timepoint, MICCAI 2021
- 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 has an interesting idea of modeling longitudinal brain dynamics with surrogate models and causal transitions, which is quite different from standard FC prediction approaches. However, the method raises some concerns around scalability, lacks comparisons with stronger longitudinal baselines, and makes restrictive modeling assumptions.
- 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 to the authors for addressing my comments. Therefore, I recommend accepting the paper. Great work!
Author Feedback
We sincerely thank the Area Chair and reviewers (R) for their constructive comments and for recognizing NDE as a “novel and promising framework” with “strong methodological contributions,” including its “causal/individualized formulation”, “rigorous validation”, and counterfactual developmental knobs as a “strong idea”. Below, we clarify the main concerns.
Comparison with longitudinal FC prediction methods (R1, R4): We agree that longitudinal FC prediction methods are relevant. However, they are not directly target-matched to NDE: existing methods primarily forecast FC matrices, whereas NDE can model developmental transitions as a causal surrogate system. In NDE, effective connectivity (EC) and counterfactual sensitivity are primary outputs, while FC only serves as a static summary of the generated dynamics. To test whether NDE’s performance can be explained by optimizing FC matching alone, we included NDE-FC, a within-framework variant trained with FC-similarity loss. Table 1 shows that although NDE-FC achieves reasonable FC prediction, it degrades EC preservation and developmental fidelity, supporting our central claim that NDE captures causal surrogate dynamics beyond conventional FC matching. We will clarify this scope distinction and discuss direct FC-forecasting comparisons as a future extension in the revision.
Assumption on residual formulation and developmental pattern (R3, R4): We would like to clarify that NDE does not impose monotonic, plateauing, linear, or additive developmental trajectories on FC edges. Eqs. 4-5 operate at fMRI signal level rather than directly on FC. Specifically, Eq. 4 adds an age-conditioned signal residual to the source-age neural surrogate to predict target-age fMRI state evolution, while FC is derived afterward from the generated time series. Thus, the residual formulation does not constrain FC trajectories. Moreover, the residual is produced by nonlinear FiLM-conditioned networks, allowing nonlinear transitions. Similarly, the age-gap gate in Eq. 5 regularizes the residual magnitude at signal level to stabilize training, without forcing FC to be monotonic or plateauing. We will clarify this distinction between signal- and FC-level constraints, while noting that the age-gap gate may reduce sensitivity to abrupt dynamical shifts.
Details on FC_free generation and developmental knob (R1): We will better define FC_free and developmental knob in the revision. As stated in Sec. 3.2, FC_free in Table 1 is computed from BOLD dynamics generated by autoregressive “free rollouts span 200 steps across four random seeds”. Regarding the developmental knob g, it is not optimized. As introduced in Sec. 2.2 through Eq. 9 and the leverage score (Imp_k) definition, g is inserted into the trained NDE with fixed parameters, and the leverage score of each region k on a targeted FC edge is computed via automatic differentiation. Top FC drivers in Fig. 3c-e are ranked by |Imp_k|, while the sign of Imp_k indicates modulation direction.
Scalability, surrogate stability, PDE fine-tuning, and statistics (R1, R4): Regarding scalability, the individualized surrogates are lightweight and trained independently, making this stage parallelizable. Although this design preserves scan-specific dynamics, we will acknowledge surrogate stability under limited timepoints as a limitation. For PDE fine-tuning, we understand the concern that limited TD samples may increase the overfitting risk. We have mitigated the risk by 1) matching scan protocol and age range between BCP and PDE data; 2) using 68 longitudinal scans from 18 TD infants; and 3) conservative fine-tuning with the identity encoder frozen and a reduced learning rate. We will clarify these and acknowledge the need for larger TD samples in the revision. Moreover, using the existing cross-validation outputs, a paired permutation test confirmed that NDE significantly outperformed competing methods across tasks (p < 0.001). We will include this in the revision.
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 reviewers agree that the paper presents a novel and promising framework for modeling longitudinal brain dynamics with strong methodological contributions, particularly in its causal and individualized formulation. However, concerns remain regarding missing comparisons to established longitudinal baselines, unclear methodological details (e.g., rollout and intervention procedures), limited statistical validation, small sample size, and unclear assumptions.
- 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 authors provides a strong rebuttal that successfully addresses the reviewers’ concerns. Overall the proposed method is novel and has uniquely interesting utility. This strength outweighs the potential drawback of missing relevant longitudinal baseline comparisons.
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.
Despite some lingering concerns, the authors have done a good job addressing the main critiques to extent allowed by MICCAI (i.e., no new experiments). This work would be of interest to the community.
