List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Logiraj Kumaralingam, Kokul Thanikasalam, Sittampalam Sotheeswaran, Jeyasuthan Mahadevan, Nagulan Ratnarajah
Abstract
Segmentation of whole-brain fiber tractography into anatomically meaningful fiber bundles is an important step for visualizing and quantitatively assessing white matter tracts. The fiber streamlines in whole-brain fiber tractography are 3D curves, and they are densely and complexly connected throughout the brain. Due to the huge volume of curves, varied connection complexity, and imaging technology limitations, whole-brain tractography segmentation is still a difficult task. In this study, a novel deep learning architecture has been proposed for segmenting whole-brain tractography into 10 major white matter bundles and the “other fibers” category. The proposed PointNet based CNN architecture takes the whole-brain fiber curves in the form of 3D raw curve points and successfully segments them using a manually created large-scale training dataset. To improve segmentation performance, the approach employs two channel-spatial attention modules. The proposed method was tested on healthy adults across the lifespan with imaging data from the ADNI project database. In terms of experimental evidence, the proposed deep learning architecture demonstrated solid performance that is better than the state-of-the-art.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_18
SharedIt: https://rdcu.be/cVD40
Link to the code repository
https://github.com/NeuroImageComputingLab/3D_Curve_CNN
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The author proposed a novel 3D deep model to classify brain white fiber tracts. Two channel-spatial attention modules are proposed and added to the backbone network architecture (PointNet) and leverage the model performance in the detection of 10 major fiber bundles.
- 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.
First, a 3D image data with high-quality labels of the major white fiber tracts is introduced in this study, which carries extensive manual efforts. This large data containing 25 individuals might benefit future studies in the related area. Second, an automatic fiber tract classification pipeline is proposed and its technic novelty sounds solid.
- 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.
-
I would doubt the name of the “segmentation model” in the framework. It is actually a classification model which categorizes a set of point curves of brain white fibers into 10 major bundles. The proposed framework requires identified fiber curves as network inputs, whose pathways are already decided and aligned in the process of whole-brain tractography. Hence, the novelty of this work is downgraded due to this setting.
-
Back to the classification task in the experiments, the comparison with baseline approaches seems unfair to me. Because the testing data are not aligned, e.g. different sizes of subjects, various quality, and even different input data formats (the most critical factor), the classification difficulty of baseline approaches are not at the same level and shall not be compared.
-
In the ablation study, I suggest the author report the averaged performance under cross-validation. We are unable to draw statistical confidence about the reliability and effectiveness of the proposed modules based on the given numbers.
-
- 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
Codes and training process are revealed in the supplementary materials. The reproducibility sounds plausible.
- 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
Please find my questions and concerns in the weaknesses section.
- 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
3
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Though there is a data contribution, the technic contribution of this study is very limited. It is basically a fiber classification model rather than a segmentation approach and the experiment settings (e.g. preprocessing and input format) lead to a relatively simple task hence bringing few impact/benefit to the area.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
5
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
2
- [Post rebuttal] Please justify your decision
Thanks for the rebuttal but I disagree with author’s points.
-
The answer to my question “it is not a segmentation model but rather a classification” is incorrect. For a segmentation model, each input file may contain many possible labels and the model will give a prediction for each unit (e.g. in pixel or voxel). However, in this study, the input is a fiber curve (points already grouped into curves) and the model only predicts a single label for this curve. This is a typical classification task which is obviously easier than the segmentation task. One of the major difficulties in segementaion is unit clustering (e.g. image pixels/voxels into unlabelled groups), which was done in the preprocess step by using the interactive dissection technique[8] and not the contribution of this study.
-
From Tab1. we can clearly find that the baseline methods use different input data formats (e.g. image data, graph data, etc.), different numbers of prediction classes (3 classes~72 classes), and even data sources (may have different resolutions, information are missing). I understand it is difficult to get public source code of baseline methods and data. However, tt is unfair to directly compare them with the proposed mode without training/testing them on the same baseline. The misaligned experimental settings could lead to a bias judgement and hence can not support the superiority of this work.
Consider all these points, I don’t think the current work gives me enough data to support their contribution.
-
Review #2
- Please describe the contribution of the paper
This work proposed a network and a new data preprocess pipeline to category the white matter tensor into 11 groups including a non-major group. The model included a T-Net and two spatial attention layers. The proposed method was validated using a manually label dataset of 25 subjects.
- 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.
As the author claimed in the introduction, the main contribution of this work is the model that can take the raw 3d curve data. It may provide an end-to-end structure and more raw information to the model. This is also partly confirmed by the results section, table 1. However, since the dataset are different, it is hard to evaluate directly.
- 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.
It may be not the weakness, but something needs to be clarified.
- what is the input of the model. The model took 256x3. I suppose that is a curve with 256 points and each point with 3d coordinates. The major fiber bundle can be selected 256 points randomly. However, did it also work on the non-major bundles, which may have less than 256 points?
- In the curve sampling section, it is not clear what is the mid points and how they were selected. Is that possible to use interpolation instead of adding points randomly to achieve the same data size?
- Ablation studies may be required to confirm the performance of the proposed model.
- Were data from different subjects mixed? Was there any difference among subjects?
- Please rate the clarity and organization of this paper
Very 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. It is not clear whether the author will distribute the manual labels.
- 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
According to the above major questions:
- Perhaps the authors can provide more details of the data sampling. especially the non-major bundle and give examples of the long vs short fibers, the single vs branched fibers.
- It would be great to compared with other methods using the same dataset.
- Ablation studies can be performed for example by removing the attention layers, or using interpolation other than randomly selected points.
- 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
It is an interesting work of classification with location and was performed on the raw data, although blind classification may be more interesting. But the performance was not fully validated and some details need to be provided.
- Number of papers in your stack
4
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #4
- Please describe the contribution of the paper
This paper proposed point cloud based 1d-CNN deep architecture to conduct whole-brain tractography segmentation. 3D raw curve points are used to represent the curves of fibers. The whole model is trained based on a classification task: the fiber are classified into 10 major fiber bundles and the “other fibers” types. The proposed method get a 98.80% average accuracy.
- 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.
Deep models cannot directly consume 3D curves in their raw data format, therefore additional preprocessing steps are needed, which adds the complexity of the segmentation task. This work uses 3D raw curve points to represent the curves of fibers and can solve the above-mentioned problem.
- 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.
-
One of the challenges of fiber segmentation is the anatomical individual variability. An effectively method should have the capability to handle the individual variances and provide robust results. In this paper, the method was evaluated using 10 subjects. However, the results part focuses on the accuracy value but the individual variability is overlooked.
-
As the accuracy values in table1 is the reported values in each work obtained using different datasets, therefore this results cannot effectively compare different methods.
-
As a major part, the motivation of Spatial attention module is not clear enough. From table2, we can see the performance of Baseline Network + Spatial attention module is worse than single Baseline Network.
-
Some sentences are not correct, such as “…processes using FSL [17] was used”, and “…which parts are the informative in …”
-
- 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 author provides code in the supplementary file.
- 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
Q5
- 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?
I have read the whole paper carefully and understand all the parts. Based on the strengths and weakness in Q4 and Q5, I give the overall score for this paper.
- Number of papers in your stack
8
- What is the ranking of this paper in your review stack?
4
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
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.
I agree with the reviewers that this is an interesting work. I would like to suggest the authors comprehensively address the concerns of all reviewers.
- 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).
10
Author Feedback
Submission 2257
We would like to thank the reviewers for their positive and critical comments and the overall accurate and careful reviews. We have already acted on most points (e.g., ablation study (Reviewer#1, Reviewer#2), data sampling (Reviewer#2), interpolation (Reviewer#2), spatial attention module (Reviewer#3), sentence correction (Reviewer#3)) in a revised, improved version of the paper, which is ready for resubmission.
There are potential misunderstandings in many of Reviewer#1’s comments that we would like to clarify here. Question 2: “First, a 3D image data with high-quality labels of the major white fiber tracts is introduced in this study” For classification, we didn’t use any image data; we used 3D raw curve points, which are detailed in the paper. All of the other comments reflect this potential misunderstanding. Question 3: “Segmentation model” We never use such a phrase in our study. The reviewer was once again confused by image-based segmentation. The goal is to segment whole-brain curves into ten major bundles and other curves. It was decided to employ the classification model for segmentation. Question 3: “pathways are already decided and aligned in the process of whole-brain tractography” We are not clear what the reviewer is referring to exactly. However, our aim is to segment the fiber curves into major bundles for a particular subject, which has not already been decided where the curve belongs to. Question 3: “testing data are not aligned, e.g. different sizes of subjects, various quality, and even different input data formats (the most critical factor)” A co-registration algorithm is used to align diffusion MR images due to variations in tract size and shape (training data and testing data). We’re confused about “different sizes of subjects.” Which size? (Images are the same size and curves are fixed size), Again, the testing data is not images. “Different input data formats”: We have used the same input data format.
Contradicting statements from Reviewer#1 “Classification pipeline is proposed and its technic novelty sounds solid” (Question 2) “the technic contribution of this study is very limited.” (Question 8)
We would like to answer the questions of Reviewer#2 in Question 3.
- “what is the input of the model…”. The input of our model is the curves with a fixed number of points of 256x3. All curves with different sizes of points were sampled to 256x3 points using our sampling technique (if the curves have less than 256 points or greater). The detailed description is now included in the revised manuscript.
- “Were data from different subjects mixed? Was there any difference among subjects?” (Reviewer#3-Question3-1) As we already mentioned clearly in the methodology part about the details of the dataset, 25 healthy subjects (males: females = 14:11, age range: 45-75) were acquired from the ADNI with the same imaging acquisition parameters. Our main aim of this work is to separate each major bundle and other curves from whole brain tractography for any of the healthy adult test subjects. Thanks to the reviewer, it is a great idea that we look into a detailed variation of different kinds of subjects, such as males and females, aging and diseases, for this problem.
Validation: Explanation (Reviewer#1, Reviewer#2, Reviewer#3) We couldn’t compare our results with related work using our data since we used our own dataset, manually created by an expert for this study. The training and testing dataset in the form of 3D curve points. And there was no related work done using the 3D curve points dataset. However, to demonstrate the usefulness of the proposed effort, we compared the performance of our technique to the reported outcomes of fiber classification models. The similar description is already in the manuscript section 3.2 Classification Results.
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
I like the idea of this paper and all the major concerns have been addressed well.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/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).
6
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The authors addressed all main concerns. I think the novel method proposed in this work is interesting and should be accepted.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/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).
5
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The key strength of this paper is to propose a novel automatic fiber tract categorization method with deep learning models. The key weaknesses are ‘unfair’ comparisons with different methods and insufficient consideration of individual variability. Although there are still some concerns, I think it is novel enough for acceptance of this paper. I would strongly suggest the authors revise the final version of this paper to integrate all useful comments.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/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).
7