Abstract

The development of a computational framework that can infer largescale brain-wide effective connectivity (EC) based on resting-state functional MRI (rs-fMRI) represents a grand challenge to computational neuroimaging. Towards the goal of estimating full-scale, whole-brain EC, we developed a new computational framework termed Large-scale nEural Model Inversion (LEMI) by utilizing a linear neural mass model with an efficient Kalman-filter based gradient descent algorithm. Key advantages of LEMI include fast estimation of both intra-regional and inter-regional connection strengths for large-scale networks, allowing exploration of both intrinsic and external mechanisms in neuroscience problems. Using ground-truth simulations, we demonstrated that LEMI can accurately and efficiently recover model parameters in a large network (100 regions) within 90 minutes. We then applied the LEMI model to an empirical rs-fMRI dataset from the ADNI database and identified widespread reduced excitation-inhibition (E-I) ratio in patients with Alzheimer’s disease (AD). Overall, LEMI provides an efficient and accurate computational framework to estimate large-scale EC and whole-brain E-I balance based on non-invasive neuroimaging data.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2060_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)

ADNI dataset: https://adni.loni.usc.edu/ HCP dataset: https://www.humanconnectome.org/study/hcp-young-adult/data-releases

BibTex

@InProceedings{LiGuo_ALargescale_MICCAI2025,
        author = { Li, Guoshi and Yap, Pew-Thian},
        title = { { A Large-scale Neural Model Inversion Framework for Effective Connectivity Estimation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {2 -- 12}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a model to estimate the effective connectivity by using linear neural mass model with Kalman-filter. The gradient descent is used to estimate the unknown parameters.

    The authors conduct both synthetic and clinical experiments using ADNI dataset. The author claims that they can accurately reconstruct the effective connectivity and show the application in ADNI for different patient groups.

  • 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 proposes a method that combines the advantages of Multiscale Neural Model Inversion and Mesoscale Individualized Neurodynamic modeling.

    It provides a relatively clear demonstration of their model and update method. The paper also validates the method by synthetic and clinical data.

  • 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. In the Fig.2, the authors verified their method in synthetic data with gaussian white noise w(t)~N(0, 0.0005) and v(t)~N(0, 0.01). Under this very limited noise, the simulation reconstruction does not look good with R^2 only above 0.6. It perhaps reveals that the proposed method is highly ill-posed. Can the author verified the method under larger noise variance?

    2. I do not understand why the author show Fig.3~5. What is the clinical significance of these weights? Did the author compare their method with popular methods in the literature and demonstrate more clearly the proposed method’s advantages? For example, some solid numerical comparison and its application. How does the linear neural mass model and Kalman-filter help the case?

    3. In the clinical experiment, we can observe some difference of the estimated weights between groups. However, what does it mean clinically? For example, can the weights be the biomarkers that can help classify patients in the ADNI? What is the classification accuracy? The weights themself does not tell any story.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The major issue lies in the experiment:

    1. The synthetic data experiment shows that the algorithm can not accurately predict the weights (with R^2 only at 0.6 level) even under very little gaussian noise. It means that the proposed method is highly ill-posed.

    2. The clinical experiment using ADNI data to estimate effective connectivity does not make sense. The author should focus more on the specific application instead of just showing the weights from estimation.

  • 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.

    After careful consideration, I think the author did not provide sufficient response to my comments. I decide to reject the paper.



Review #2

  • Please describe the contribution of the paper

    The author developed the Large-scale nEural Model Inversion (LEMI) framework to overcome limitations in current macroscale network effective connectivity and to identify both intra-regional and inter-regional EC.

  • 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 well-written. The framework has been validated on both synthetic data and real-world resting-state fMRI, yielding promising results.

  • 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.

    Please see the comments 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 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

    Major comments:

    1. Limited Synthetic Data: The synthetic data is quite limited, as only TR = 1 s and a canonical hemodynamic response function (HRF) are simulated. However, the BOLD signal can exhibit more complexity beyond a single assumption. Consider exploring alternative HRF models or justifying this choice. References: -Buxton, R. B., et al. Modeling the hemodynamic response to brain activation NeuroImage 23, S220–S233 (2004). -Chuang, K.-C., et al. Joint Estimation of Neural Events and Hemodynamic Response Functions from Task fMRI via Convolutional Neural Networks. International Workshop on Machine Learning in Clinical Neuroimaging. Cham: Springer Nature Switzerland, 2023.
    2. Choice of Wiener Deconvolution: The rationale behind selecting Wiener deconvolution for estimating neural activity from the BOLD signal is unclear. Please justify why this approach was chosen over other possible methods.
    3. Figure 2 – Summary of Results: While it is understandable that only a subset of results is presented due to page limits, it would be beneficial to include a summary of: (a) The accuracy of E neural activity estimation (b) The prediction performance of the deconvolved BOLD signal across different ROIs.

    Minor comments:

    1. Figure 3 Title: The title of Figure 3 is unclear. Please explicitly state what the correlation coefficients represent. Specifically, clarify that the figure shows the distribution of correlation coefficients between estimated and ground-truth parameters for all 30 synthetic subjects.
  • 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?

    Positive results have been presented; however, the limited scope of the synthetic data and the vague processing of the BOLD signal weaken the robustness of the paper.

  • 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.

    The authors addressed my major comments. The updates and revisions can be feasibly implemented in the final version of the paper.



Review #3

  • Please describe the contribution of the paper

    The main contribution is a method that estimates effective connectivity (EC): 1. among a relatively large number of brain regions (roughly 100), 2. with modeling of both excitatory (E) and inhibitory (I) neural populations within each region. Prior methods were able to do 1 or 2 but not both together.

  • 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 main strength is the ability to include larger numbers of regions in EC analysis, while also modeling multiple neural behaviors within one region. I think we all want full-brain EC analysis at a very fine scale with few constraints placed on the EC modeling, but we know that this is too computationally hard at this time. This paper tries to move us toward that goal which is laudable.

  • 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.

    Perhaps the main weakness is that this EC model makes a variety of simplifying assumptions that constrain what types of EC relationships can be detected, but the set of such assumptions is not clearly laid out, nor what the repercussions of those simplifying assumptions are. For example, the model estimates neural activity making a variety of assumptions. There are simplifying assumptions about the hemodynamics associated with neural activity. There is the classic “2/3 of activity is E, 1/3 is I” assumption. Kalman filtering makes its own assumptions about linearity of relationship between state and prediction. It is not clear how unrealistic the resulting modeling is after making all these assumptions.

    Another weakness is that this method is supposed to be combining the strengths of MNMI and MINDy, but the comparator method in the experiments is neither MNMI nor MINDy. Instead the comparator is some other method— rDCM— that assumedly is inferior. So it is difficult to tell what value is added by the proposed method over what already exists in the literature.

    There are experiments from purely synthetic data generated using the equations that govern the model; and there are face validity experiments using real data. Those are fine, but the study lacks experiments using data output by a BOLD simulator such as STANCE. The advance of simulator data is that on one hand the true / correct result is known; and also the data bears some resemblance to real BOLD data.

    The experiments make a set of seemingly arbitrary design decisions that leave room for cherry-picking data that is more favorable to the algorithm. They picked 48, 48, and 48 NC, MCI, and AD participants from ADNI, even though the number of people in each category is much larger. How and why did they do this picking? They picked ROIs from 4 functional networks to analyze even though the number of possible networks is larger. How and why did they do that picking? They got the structural network for the ADNI participants, from an unrelated young adult data set, HCP— why, and how do we know this structural network accurately depicts elderly brain connections? They selected the top 10% of the structural connections for analysis— why 10%? These all seem to be arbitrary parameter settings without justification.

  • 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?

    I generally think that whole-brain EC is an important goal for the field. But I am not sure about the validity of the experiments, nor the limitations of the methodology, for the reasons listed above.

  • 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 responses to critique are pretty reasonable, and while the paper does have limitations I believe the larger goal of pushing toward brain-wide EC estimation is being served in a substantial way.




Author Feedback

Reviewer #1 has concerns about limited synthetic data and the choice of Wiener Deconvolution. We chose the canonical HRF because (1) it is the basis function whose parameters best fit the BOLD response observed by the majority of empirical studies; and (2) it has been used by many fMRI models with proven accuracy (PMID: 28259780, PMID: 32603858). We plan to use a variable HRF in future improvement of LEMI. Also, we have tested different TR (1-3 s) and found the results are robust. We chose Wiener Deconvolution due to its optimal noise reduction and improved signal restoration, which has been widely used in signal and image applications.

Reviewer #1 has an additional critique about the lack of a summary of results. This is an excellent suggestion. We already have the relevant results and can readily include them in the final version of the paper.

Reviewer #2 raises several fair points about the model assumptions. We fully recognize these assumptions. At the same time, we want to emphasize that these assumptions are biologically plausible and are necessary to enable large-scale EC computation. For example, the relative contribution of excitatory and inhibitory neural activity to the BOLD signal, approximately two-thirds excitatory to one-third inhibitory, is consistent with previous fMRI models (PMID: 26335064; PMID: 34284335), reflecting the higher proportion of excitatory neurons in the brain. We intend to make these model assumptions explicit in the final version of the paper and investigate their impact in a future study.

Reviewer #2 is concerned with our choice of the comparison method (i.e., use of rDCM instead of MINDy or MNMI). Based on our previous study, EC estimated by rDCM is highly correlated with EC estimated by MINDy, so these two approaches generate qualitatively similar results. We also compared LEMI with MNMI and found that while the results are consistent with each other, LEMI runs much faster than MNMI (tens of mins vs. hours).

Reviewer #2 has concerns about the study design. We used the fMRI data from a previous study as they are readily available. The fMRI dataset was taken from ADNI-GO/2 studies with 48 AD subjects. We selected the same number of age- and gender- matched NC and MCI subjects. We chose the default mode, salience, frontoparietal control and limbic networks because we want to focus on the networks that are most implicated in AD pathology. We agree with the reviewer that it may be more realistic to use a BOLD simulator - this will add great value to further validating our method.

Reviewer #2 has additional concerns about the use of HCP structural networks and the percentage of structural connections. As DTI data is not available for every ADNI subject, we used high-quality DTI data from HCP. The impact of this limitation should be minimal since we used the average SC from 100 HCP subjects for SC selection. We have varied the percentage of connections from 10% to 30% and found the results remained qualitatively the same.

Reviewer #4 has concerns about the limited noise. In the synthetic simulation, the standard deviation (SD) of the neural state and measurement is around 0.2 (see Fig. 2A, B). With a variance of 0.005 (not 0.0005) and 0.01, the SD of the noise is 0.07 and 0.1. So the noise-to-signal ratio is about 35%. Thus, the noise level is not low. We have tested higher noise level (SD up to 0.3) and the results remain unchanged. Also, a R^2 of 0.6 is considered a good estimate (PMID: 32603858) with p values close to 0.

Reviewer #4 has additional concerns about the utility of the clinical experiment. We respectfully disagree that the experiment has little value. Figs. 4 and 5 show the excitation level in MCI/AD is significantly reduced compared to NC. These results are biologically important, indicating disrupted E-I balance in MCI/AD which can be used as biomarkers. Due to the page limit, we could not include the classification results but would like to follow up in a future study.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top