Deconvolving Complex Neuronal Networks into Interpretable Task-Specific Connectomes

Summary

This paper addresses the challenge of identifying elementary functional neuronal networks and their combinations in the context of complex tasks, using task-specific functional MRI (fMRI) data. The central problem it tackles is the deconvolution of task-specific aggregate neuronal networks into elementary networks. These elementary networks can then be used for functional characterization and mapped to underlying physiological regions of the brain. Due to the high-dimensionality, small sample size, acquisition variability, and noise inherent in this task, the authors propose a deconvolution method based on supervised non-negative matrix factorization (SupNMF). The results demonstrate that SupNMF can uncover cognitive "building blocks" of task connectomes that are physiologically interpretable, predict tasks with high accuracy, and outperform other supervised factoring techniques in both prediction accuracy and interpretability. Overall, the proposed framework offers valuable insights into the physiological foundations of brain function and individual performance.

Strengths

1. The paper presents a valuable effort to implement a supervised decomposition method in a novel way, showing the potential for this approach in a complex context, such as the analysis of neuronal networks. 2. The authors provide fascinating results indicating that each task has unique markers within these learnable networks. This insight could contribute significantly to understanding how tasks are represented and processed within the brain. Observing that some networks are shared across tasks also provides a meaningful direction for future research. 3. The alignment of the findings with existing physiological research is a good sense for future study.

Weaknesses

1. Reproducibility: The study could be enhanced by applying the proposed method to other datasets or by resampling the existing dataset. This would help to assess the generalizability of the method and the robustness of the findings, which is currently a limitation of the work. 2. Baseline Comparison: It would be beneficial if the authors had compared the proposed SupNMF method with other supervised decomposition methods, such as Partial Least Squares regression. This lack of comparison limits the understanding of how their proposed method stands in relation to existing methodologies regarding performance and effectiveness. 3. Parameter Study: The authors need to provide an in-depth analysis or sensitivity study concerning the weight parameter \lambda. As this parameter likely plays a significant role in balancing different loss terms, this omission constitutes a substantial weakness, potentially leaving readers unclear about the effectiveness of supervision signals.

Questions

1. In the abstract, line 9 uses the wrong left double quotation mark and right double quotation mark. 2. Provide a reference and explanation about UMAP mentioned in the paper. For the benefit of readers who might not be familiar with this method, I recommend that the authors reference a key source on UMAP and provide a brief explanation of its use and significance in this context. 3. In the related work, there is a missing related paper when talking about the interpretable GNN, "Interpretable Graph Neural Networks for Connectome-Based Brain Disorder Analysis. MICCAI 2022"

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Yes

Summary

This paper presents a decomposition method for task-functional connectivity. It proposes canonical task connectomes which derives sub-structure of functional brain connectivity which group connectomes which identify elementary components of the overall connection. The authors use supervised non-negative matrix factorization to factor connectome matrices and show that the derived features are suitable to predict tasks for functional MRI and robust dimension reduction of the original representation.

Strengths

- The paper is clearly written. - It construct a clear optimization problem whose results are easily interpretable. - It demonstrates a solid dimension reduction for connectome data.

Weaknesses

- Motivation for including supervision in the connectome decomposition is very weak. - There are some missing details on variable descriptions, e.g., $d$ in line 96 is missing, $\hat y$ in eq (2) is missing (although can be inferred that it is a prediction for a class), and derivation of $X$ from $C$ should be better explained. - Experiment is performed only on one study. It can be excused if the dataset is rare, but there are so many publicly available fMRI data. - Lack of baselines. It is missing the most fundamental baseline, i.e., LDA. Moreover, just typing in "supervised dimension reduction" in google scholar yields various literature but this paper demonstrates only SupSVD as a supervised baseline.

Questions

- I am not sure what the role of supervised SVD is in this paper. It is written as a separate section in this paper but is used as a baseline in the experiment. I think section 2.3 can be removed and filled by other details of the proposed method. - Utilization of other datasets? The authors mention several public neuroimaging datasets in the introduction but the proposed framework is validated only on a single benchmark. I believe validating the method on other neuroimaging studies will strengthen the paper. - It is quite straight forward that task-wise supervision during decomposition will, of course, increase the accuracy of the downstream task prediction. Are there other benefits? What if there is a label set difference between the training and testing set? - Including supervision in dimension reduction / decomposition has a long history. I think the very very basic baseline should be LDA rather than NMF or SVD as it is a supervised method. - Perhaps the authors should discuss why SupNMF is outperforming SupSVD in Fig 2.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The paper does not discuss any limitation of its own.

Summary

The paper presents a new approach to identify task-specific building blocks of neuronal activity from fMRI data by using supervised matrix factorisation. The identified patterns generalise from the train to a test set and match expectations on the brain activity for the different tasks from the neuroscience literature.

Strengths

The paper is well written and addresses an important problem in the analysis of fMRI data in a novel way that achieves impressive performance.

Weaknesses

While the paper criticises that existing methods cannot be applied to large, diverse datasets, the paper lacks a study of the computational efficiency of the proposed approach and a comparison with existing methods.

Questions

- How well does the method scale? - What results do existing methods achieve in the performed experiments? e.g. ICA-based [13, 10, 34] or other ML-based methods [19, 26, 33, 29]

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations of the approach are not openly discussed.

Summary

This contribution presents a novel method to find a functional basis for a database of task fMRI acquired from different subjects. The functional basis, dubbed canonical task connectomes, is shard across large cohorts; can be composed into task-specific networks; and is predictive of task efficacy. The authors produce this functional basis through supervised and non-supervised NMF and SVD. For this they propose an objective function (in equation 3) which is compatible with these methodologies. To implement this approach the authors use the HCP100 database which has 6 cognitive tasks. To show that the obtained basis is task-specific the authors use UMAP plots of different resulting decompositions showing good separability of tasks in the embedded UMAP space. To show that their basis is generalizable across cohorts they use 80/20 splits of the HCP 100 database and use the basis to inform a classifier that predicts the task performed by an unseen subject given the fMRI acquisition. To claim physiological and anatomical grounding for their basis, the authors compare qualitatively the basis against known anatomical traits and the involvement of different basis components as important features for each task.

Strengths

This contribution is a high-quality application of known methods to the significant problem of understanding how functional connectivity in the human brain (as measured by fMRI) is centric to cognitive tasks. The manuscript presents a well-justified formulation of the problem as a deconvolution case and solves it through different approaches, supervised and non-supervised. This formulation and resolution are original and well presented. Even if the methodological contribution is not at the center of this manuscript, the application of known techniques is well-justified and evaluated. The evaluation of the results are a good balance of qualitative evaluation (e.g. with UMAP embeddings), and quantitative (e.g. with the clustering approaches) in the case of task-specificity of the connectomes. The generalisation experiment using a downstream classification task is also well conceived. Finally the relation with anatomy and physiology is well organised. In all, this contribution presents a very good application of known methods to an important problem in neuroimaging. So it's a paper that will have impact in one area, the neuroimaging one.

Weaknesses

I find two weaknesses which are related to claims of cohort generalisability. In short, with the availability of public datasets of fMRI, it's hard to justify a cohort generalisation claim while staying in one 100-subject database. Specifically when the used 100-subject set is a subsample of a 1,200 total database. In light of this, second weakness I find is the lack of an analysis of the stability of the functional basis across datasets, including a study on dataset size.

Questions

My main questions will be database specific. First, how many subjects are needed to obtain the functional basis, or canonical task connectomes? For this the authors could provide a learning curve-style analysis analysing the dispersion of the found connectomes with respect to the sample size. Second, to properly claim cross-cohort the authors should properly use a second large cohort, such as UK Biobank or ABCD which, admittedly, have different task fMRI protocols. Short of this, the authors might look into toning done the cross-cohort claim.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The authors have not explicitly mentioned the limitations in the manuscript.

Summary

This paper presents a novel framework for fMRI analysis that aims to deconvolve complex neuronal networks into task-specific elementary networks called "canonical task connectomes." The proposed method utilizes supervised matrix factorization to identify these task-specific networks and demonstrates their interpretability and generalizability. The study showcases experimental results on the Human Connectome Project dataset, highlighting the ability of the framework to capture the natural task-specific structure in neuroimages.

Strengths

- The paper presents a new problem formulation and introduces the SupNMF method, which is a novel approach to identifying task-specific networks. The authors demonstrate the usefulness of the proposed framework in identifying canonical task connectomes that have a strong physiological basis and can be mapped to regions of the brain to identify physiological underpinnings of tasks. - the authors present the problem formulation and the proposed method in a clear and concise manner. - The proposed interpretable framework has the potential to advance understanding of complex cognitive processes and to identify biomarkers for predicting tasks. - The authors also provide a comprehensive discussion of relevant methods and materials.

Weaknesses

- While the authors present comprehensive experimental results, they could provide more details on the performance of the proposed framework in comparison to other state-of-the-art methods. Additionally, the authors could provide more details on the interpretability of the identified canonical task connectomes and how they relate to existing literature in neurosciences. - While the authors briefly mention the potential applications of the framework in understanding shared and unique functional networks across different pathologies and how task-specific networks can get dysregulated due to the onset and progression of diseases, a more in-depth discussion of these applications and their potential impact on the field would be helpful. - The authors did not explain much on why the “unrelated set” of subjects in the Human Connectome Project is selected. Also, more datasets are expected to be included to demonstrate the generalizability of the proposed method. - The authors could provide more details on how they determined the optimal number of latent connectomes and how this choice impacts the results. One potential concern of the proposed framework is that it relies on the assumption that the observed connectome matrix can be represented as a linear combination of a small number of latent matrices. While this assumption may hold for some datasets, it may not be applicable to all fMRI datasets, especially those with high levels of noise or variability. Additionally, the choice of the number of latent connectomes (i.e., the dimensionality of latent space) is critical and may impact the performance of the proposed framework. Another potential weakness of the proposed framework is that it requires task-label vectors for each connectome in the dataset. While the authors provide details on how they obtained the task-label vectors for the HCP dataset, it may not be feasible to obtain such labels for all fMRI datasets. Additionally, the choice of the task-label vectors may impact the performance of the proposed framework, and the authors could provide more details on how they selected the task-label vectors and how this choice impacts the results.

Questions

- Could you provide a more detailed comparison of the performance of the proposed framework with more state-of-the-art methods of other types (CNNs, GNNs)? How does the proposed framework outperform or differ from existing approaches in terms of accuracy and generalizability? - Can you elaborate more on the interpretability of the identified canonical task connectomes? How do these connectomes relate to existing literature in neurosciences? Are there any specific brain regions or networks that are consistently identified across different tasks? - How generalizable are the findings of this study? Are there any potential biases or confounding factors that could impact the results? - How does the proposed framework handle potential confounding factors such as motion artifacts or physiological noise? Were any specific preprocessing steps or techniques employed to address these confounds?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

While the proposed framework has the potential to advance our understanding of complex cognitive processes and to identify biomarkers for predicting tasks, it is important to consider the potential ethical implications of this research. For example, the use of fMRI data for predicting cognitive states or identifying biomarkers could raise concerns about privacy, informed consent, and potential misuse of the data.