List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Naresh Nandakumar, Komal Manzoor, Shruti Agarwal, Haris I. Sair, Archana Venkataraman
Abstract
Parcellations used in resting-state fMRI (rs-fMRI) analyses are derived from group-level information, and thus ignore both subject-level functional differences and the downstream task. In this paper, we introduce RefineNet, a Bayesian-inspired deep network architecture that adjusts region boundaries based on individual functional connectivity profiles. RefineNet uses an iterative voxel reassignment procedure that considers neighborhood information while balancing temporal coherence of the refined parcellation. We validate RefineNet on rs-fMRI data from three different datasets, each one geared towards a different predictive task: (1) cognitive fluid intelligence prediction using the HCP dataset (regression), (2) autism versus control diagnosis using the ABIDE II dataset (classification), and (3) language localization using an rs-fMRI brain tumor dataset (segmentation). We demonstrate that RefineNet improves the performance of existing deep networks from the literature on each of these tasks. We also show that RefineNet produces anatomically meaningful subject-level parcellations with higher temporal coherence.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_30
SharedIt: https://rdcu.be/cVD5c
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, the authors introduce RefineNet, a Bayesian-inspired deep network architecture that adjusts region boundaries based on individual functional connectivity profiles. RefineNet uses an iterative voxel reassignment procedure that considers neighborhood information while balancing temporal coherence of the refined parcellation.
- Please list the main strengths of the paper; you should write about 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.
They present RefineNet, a new method to optimize the functional brain parcellations based on the existing parcellations, wihch is a a Bayesian-inspired deep network architecture.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
The architecture of RefineNet was not cleary described. And the authors stated that they obtained the optimized functional brain parcellations with higher temporal coherence, in order to validate it, they plotted parcellation cohesion in Fig.3, but why is the cohesion of RefineNet than that of Combined?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The reproducibility of the paper seems not very high, the RefineNet architecture is not very detailed and clear.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
In addition to the above problems, the author should rearrange the sturcture and content of the paper, especially in method and result sections. there is only a Section 3.1 in Section 3, and acctually there are various results in section3.1. The authors should rehearsal expression and description, i.e., in “The coherence term S uses the pearson correlation coefficient with each mean time series”, Whose and whose relevance is this?
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper provided a method to optimize the functional brain parcellations based on the existing parcellations according to functional profiles, wihch can be appended to existing networks.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
3
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
The authors developed a novel network architecture that can update parcellation scheme based on resting state information.
- Please list the main strengths of the paper; you should write about 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 network is flexible that it can be used solely to derive the best parcellation for resting state data, or can be used for joint training to optimize some task performance
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
none
- Please rate the clarity and organization of this paper
Excellent
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
no concern
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
The paper is well written. the results are convincing, particularly when trained for tasks, the cohesion measure for resting state is still higher using the proposed method than the original one.
Maybe one super minor comment: the RefineNet by itself is very shallow and may not be called “deep” neural network.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
8
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
method is novel. writing is clear. results are convincing.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
This paper proposed a framework for task-specific individualized functional brain parcellation.
- Please list the main strengths of the paper; you should write about 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 framework proposed in this paper integrates the prediction task with the process of individualized parcellation learning, which may benefit to understand the different brain functional organization contributes different prediction task.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1) The clarity of the paper, especial the method section, could be improved. 2) The method to select the best hyper-parameter configuration are not clearly described. 3) It is not clear the improved accuracy is from the integration of the prediction task with the process of individualized parcellation learning. It seems like the RefineNet-only model cannot consistently improve the performance comparing with group-level parcellation that most of the individualized parcellations claim to outperform. Then, it is necessary to obtain the brain parcellation with the state-of-the-art individualized parcellation method and adopt the corresponding deep learning framework used in the three tasks. Then compare the results with the “Combined” proposed in this paper.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
Good.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
1) Please improve the clarity of this paper, especially the method section. 2) The framework of the proposed method illustrated in Fig. 1 can be improved. Some key words or symbols in Fig. 1, e.g., network weights, ne(v),A_v^j,… are not explained clearly in the main content. It is better to place the definition of s_{v,p} after mu_1,…mu_P and near S. 3) It looks like the experiments did not include all the subject in the respective datasets, please explain the subject exclusive criteria clearly to improve reproducibility. 4) It is better to describe the method of how to select the best hyper-parameters in the proposed method. For example, how to determine that epoch= 5 and I=20. 5) To validate the effectiveness of the proposed method, is it possible to design another comparison to make sure the better accuracy comes from the co-training strategy? In the comparison, first, obtain the brain parcellation with the state-of-the-art individualized parcellation method; then adopt the corresponding deep learning framework used in the three tasks to get the prediction results; At last, compare the results with the “Combined” proposed in this paper. 6) An extra “a” in the first paragraph of section 2.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes a method for task-based individualized parcellation, which is different with the traditional parcellation framework.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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.
This paper proposed a RefineNet framework to perform task-specific individualized functional brain parcellation. The key strength is to integrate prediction task with individualized brain parcellation. However, there are some weaknesses: 1. Unclear description of method part, e.g., the architecture of RefineNet, selection of hyper-parameter configuration, subject exclusion criteria, etc.; 2. The organization of this paper should be improved (e.g., there is only Section 3.1 in Section 3). 3. Limitation of comparison with SOTA methods. All the reviewers and meta-reviewer acknowledge the certain novelty of this study for benefiting the understanding of contributions of different brain functional organization to different prediction tasks. The authors should address all reviewer comments into the final paper.
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
2
Author Feedback
We thank the reviewers for their careful evaluation of our work. Below, we clarify misunderstandings about our methodology and describe minor changes to our paper.
Reviewer 1: The reviewer is correct that the temporal cohesion of the RefineNet Only model is higher than that of the combined model. This result is intuitive, as reassignment from the RefineNet only model is driven entirely by the temporal coherence loss term. In contrast, the combined model must balance temporal cohesion with the downstream task performance. The experimental results are meant to be taken together, where the combined model outperforms both the original and RefineNet Only models in task performance while still maintaining a significantly higher parcel cohesion than the original model. With regards to the paper structure, we have created a Section 3.1 and Section 3.2 in Experimental Results. Finally, we disagree with the reviewer about the model presentation, as we have conscientiously tried to describe each modeling component. If the reviewer has specific suggestions for clarification, we are happy to make these changes to our paper.
Reviewer 2: We agree that RefineNet is a shallow model and will remove the term “deep” in reference to our architecture throughout the paper.
Reviewer 3: With regards to paper clarity, we will modify Fig. 1 as suggested. The reviewer also brings up a good point about the ambiguity of our hyper-parameter selection. The hyper-parameter I was set to be 20 because we empirically observed that this was large enough to provide robust reassignment in the RefineNet Only model while maintaining a reasonable training time. We use cross validation on the regression task with a separate 100 HCP subjects to determine e=5 for alternating training. Both lambda and e were constant across all networks and datasets to minimize data leakage. We apologize for omitting these details in our paper and will include them in our final version.
Regarding the comment about subject inclusion, we randomly sub-selected 300 subjects from the HCP dataset for the regression task to have a dataset comparable in size to that of the original M-GCN paper. The cohort used in the M-GCN paper included 275 HCP subjects. We obtained our ABIDE II subjects as a part of a larger study on multimodal connectivity. The 233 subjects are selected based on having both diffusion MRI and rs-fMRI available.
Reviewer 3 asks if it is possible to compare RefineNet with a state-of-the-art method to ensure that the improved performance comes from our co-training strategy. To the best of our knowledge, RefineNet is the only parcellation refinement method that can be amended or attached to existing models and trained jointly. Thus, we argue that there does not exist a suitable parcellation refinement method for comparison. We acknowledge two prior works that perform subject-specific refinement in the Introduction of our paper: (1) Nandakumar et al., “Defining Patient Specific Functional Parcellations in Lesional Cohorts via Markov Random Fields” and (2) Chong et al. “Individual Parcellation of Resting fMRI with a Group Functional Connectivity Prior”. However, these methods cannot be trained in a task-aware fashion, and empirically, we find that they tend to randomly collapse parcels across patients at higher resolutions. Hence, we opted not to include these methods in our paper.