Paper Info Reviews Meta-Review Author Feedback Post-rebuttal Meta-Reviews

Authors

Yuqian Chen, Chaoyi Zhang, Yang Song, Nikos Makris, Yogesh Rathi, Weidong Cai, Fan Zhang, Lauren J. O’Donnell

Abstract

White matter fiber clustering (WMFC) enables parcellation of white matter tractography for applications such as disease classification and anatomical tract segmentation. However, the lack of ground truth and the ambiguity of fiber data (the points along a fiber can equivalently be represented in forward or reverse order) pose challenges to this task. We propose a novel WMFC framework based on unsupervised deep learning. We solve the unsupervised clustering problem as a self-supervised learning task. Specifically, we use a convolutional neural network to learn embeddings of input fibers, using pairwise fiber distances as pseudo annotations. This enables WMFC that is insensitive to fiber point ordering. In addition, anatomical coherence of fiber clusters is improved by incorporating brain anatomical segmentation data. The proposed framework enables outlier removal in a natural way by rejecting fibers with low cluster assignment probability. We train and evaluate our method using 200 datasets from the Human Connectome Project. Results demonstrate superior performance and efficiency of the proposed approach.

Link to paper

DOI: https://doi.org/10.1007/978-3-030-87234-2_47

SharedIt: https://rdcu.be/cyl8J

Link to the code repository

N/A

Link to the dataset(s)

https://db.humanconnectome.org/

Reviews

Review #1

Please describe the contribution of the paper

The method presented shows a unsupervised white matter clustering. They solve the problem by using CNN to learn embeddings of the input fibers. In addition, the authors used the anatomical segmentation data for learning the embeddings of the input data. The authors used 200 data from the Human Connectome Project. The authors use modified Deep Convolutional Embedding Clustering (DCEC) which is used for image clustering task.
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 authors present a great literature review presenting most of the important works in white matter clustering. In addition, the authors have opted to use the unsupervised clustering for white matter matter clustering. This is particularly thoughtful as clustering white matter labels is a very labor intensive task. In addition, often there are disagreements on the existence of a particular fiber tract. In such circumstances, taking a direction where the algorithm learns the natural embedding of the fibers is pretty interesting and smart. The results shown in Figure 2 are pretty convincing and excellent.
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.

Authors do not quite agree on a single pairwise metric. Refer to FiberNet 2.0, where authors have discussed handling noisy fibers (line 6 on page 2). As authors have discarded fibers shorter than 40 mm, they should also discard extremely long fibers. Sometimes, there are extremely long fibers that are produced by chance but do not have any anatomical significance. Another important point is that in the HCP datasets, all the images are resampled to a common space, but in the real world all the images might not be normalized to the same space. So the diffusion images need to be resampled in the MNI space or the tracts need to be transformed in the MNI space. The authors should address these issues.
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

I wasn’t clear on the input format of the data. However, as the authors have showed interest in releasing the code. I will assume that with a good example set, these details will be more 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://miccai2021.org/en/REVIEWER-GUIDELINES.html

I think the overall paper is quite well written and I like the unsupervised clustering techniques for white matter clustering as labelling white matter tracts is very expensive and time consuming. However, I would like the authors to address if the outcome of the algorithm changes if the distance metric is changed. There is no fixed distance metric used in defining distances between the curves as mentioned in reference 10 or related references.
Please state your overall opinion of the paper

strong accept (9)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

I think the unsupervised clustering method employed by the authors is very interesting and is a key to white matter clustering. The paper is well written and the details are properly explained. I had some minor points which might improve the paper and they are mentioned in the review.
What is the ranking of this paper in your review stack?

1
Number of papers in your stack

5
Reviewer confidence

Very confident

Review #2

Please describe the contribution of the paper

The authors propose an unsupervised deep learning-based framework for white matter fiber clustering (WMFC). In addition, brain anatomical segmentation data is used to assure the anatomical coherence of the fiber clusters. The proposed method is trained and evaluated using 200 datasets (100 training, 50 validation and 50 testing) from the Human Connectome Project.

The proposed framework is compared with two state-of-the-art WMFC methods, WhiteMatterAnalysis (WMA) and QuickBundles (QB).

The paper appears well written and of general interest.
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 model is the first of its kind proposed using self-supervised learning.

Brain anatomical segmentation data is used to assure the anatomical coherence of the fiber clusters. This is an important step in the removal of outlier fibers.
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 well-known disadvantages of unsupervised learning is that the spectral properties of classes can change over time therefore the same class information cannot be ensured while moving from one image to another. Can the authors please explain if this was considered as a limitation? And if yes, how was it addressed?

Stringent outlier removal can potentially eliminate true positive fibres that deviate from the normal tract trajectory. How did the authors make sure this was avoided?
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

Based on the information provided by the authors upon submission the proposed work appears reproducible at this stage.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html

How long did it take to interpret and label the classes following the clustering in training?

Can the authors please provide more details on data pre-processing? Especially for the anatomical information incorporated into the neural network to improve cluster anatomical coherence.

Page 3: “our approach has demonstrated superior performance and efficiency via evaluations on a large scale dataset” – superior performance when compared to what? Can the authors please rephrase?

I recommend the authors to clearly delineate the two stages of pre-training and clustering in Figure 1. I believe this would be of added value for the reader. Page 5: “We evaluated our method on 200 healthy adult datasets from the Human Connectome Project” – The authors have used 100 datasets to train the proposed method and evaluated the method on a subsequent 100 datasets. I recommend the authors to correctly state the above in the manuscript.

How reusable is this model? Can the authors please provide more details on the reproducibility of the model using independent datasets? If this is not yet addressed, can the authors specify it as a limitation of the study?
Please state your overall opinion of the paper

accept (8)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

I believe the proposed work is in line with the MICCAI areas of interest and has academic merit.
What is the ranking of this paper in your review stack?

3
Number of papers in your stack

4
Reviewer confidence

Confident but not absolutely certain

Review #3

Please describe the contribution of the paper

This article proposes an unsupervised deep clustering algorithm for white matter tractograms based on the Deep Embedding Clustering (DEC) framework [33]. The main novelty lies in the definition of the latent space where the clustering is performed. Authors propose to use a self-supervised strategy where the distance between two encoded fibers in the latent space should be similar to the actual distance between the two fibers in the brain space. Authors also propose to adapt the DEC framework for taking into account anatomical 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 article is clear and easy to read
- It successfully adapts a SOTA deep clustering algorithm for white matter clustering
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.
- Not clear the clinical usefulness. Are these clusters clinically relevant and reproducible?
- The evaluation is limited to only two existing algorithms and to a precise choice of hyper-parameters (i.e. number of points per fiber and number of clusters)
- Some notations and concepts are not clear. For instance, it is not clear which distance authors use. Is it the minimum average direct-flip (MDF) distance as in [7]?
- Some choices are not well discussed. For instance, why simply looking at pair of distances and not using triplet loss or contrastive learning? In Sec.2.3, if the Dice score D_{ij} is zero with all clusters, the soft assignment label is not actually zero, is it desired? Why?
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

Authors don’t share their code (and don’t’ say if they will do after review) but they give enough details to reproduce their experiments.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html
- I have concerns about the clinical usefulness of the estimated clusters. Can they be used by clinicians? Are they reproducible among subjects and consistent with anatomical tracts? This is not demonstrated nor discussed. Why should one use the proposed algorithm and not WMQL, Classifyber or TractSeg, just to mention a few?
- The definition of the embedding space is not sufficiently discussed, and some choices are not well motivated. Why not using the more common triplet loss or contrastive losses? Let A and B be two fibers and C the inverse of A. With the proposed framework, the distance between A and B will be the same as the distance between C and B and this will be mimicked in the embedding space. However, there is no constraint that forces A and C to be mapped to the same point in the embedding space. Is it desired? Why? More explanations are needed.
- When incorporating the anatomical information, authors introduce a dice score (i.e. a positive value between 0 and 1) which is multiplied by the distance to the relative cluster. It seems therefore that if a fiber is quite far from a cluster (high distance), the corresponding Dice score is likely to be around zero and therefore also their product. Is it desired? If yes, why? More explanations are needed.
Please state your overall opinion of the paper

probably reject (4)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

Please see previous sections.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

3
Reviewer confidence

Confident but not absolutely certain

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.

The authors propose an unsupervised deep learning-based framework for white matter fiber clustering (WMFC) and compare it with 2 other state of art techniques. The work is novel and should be accepted.
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).

1

Author Feedback

We thank the reviewers for their helpful comments. As requested, we will clarify several concepts in the final version of the paper, including the input data and preprocessing, the tractography registration, the Dice loss, and the fiber distance metric definition. We note that our network is designed to address the pretext task of distance prediction in self-supervised learning. This is related to contrastive learning (as mentioned by a reviewer), where contrastive learning relies on pairs of similar and dissimilar examples rather than a distance between examples as in our work. In the final paper, we will clarify that the results of the experiments in 100 HCP datasets strongly demonstrate the reproducibility of the clusters across subjects. Finally, we will clarify that fiber clustering has been shown to be useful in multiple applications including atlasing and labeling of fiber tracts, as well as machine learning and statistical analyses of the white matter.

back to top

Deep Fiber Clustering: Anatomically Informed Unsupervised Deep Learning for Fast and Effective White Matter Parcellation