List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Geng Chen, Haotian Jiang, Jiannan Liu, Jiquan Ma, Hui Cui, Yong Xia, Pew-Thian Yap
Abstract
Advanced contemporary diffusion models for tissue microstructure often require diffusion MRI (DMRI) data with sufficiently dense sampling in the diffusion wavevector space for reliable model fitting, which might not always be feasible in practice. A potential remedy to this problem is by using deep learning techniques to predict high-quality diffusion microstructural indices from sparsely sampled data. However, existing methods are either agnostic to the data geometry in the diffusion wavevector space (q-space) or limited to leveraging information from only local neighborhoods in the physical coordinate space (x-space). Here, we propose a hybrid graph transformer (HGT) to explicitly consider the q-space geometric structure with a graph neural network (GNN) and make full use of spatial information with a novel residual dense transformer (RDT). The RDT consists of multiple densely connected transformer layers and a residual connection to facilitate model training. Extensive experiments on the data from the Human Connectome Project (HCP) demonstrate that our method significantly improves the quality of microstructural estimations over existing state-of-the-art methods.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_11
SharedIt: https://rdcu.be/cVD4S
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
This paper introduced a hybrid graph transformer (HGT) method to estimate tissue microstructure measures by using a graph neural network (GNN) for q-space modeling and a residual dense transformer (RDT) for spatial modeling. The proposed method was compared with several state-of-the-art methods and showed better performance for estimating NODDI parameters with reduced samples. Moreover, an extensive ablation study was performed to examine the effectiveness of the learning modules.
- 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.
1) The HGT method includes modules for both q-space and x-space learning that is very suitable for modeling diffusion MRI and different from previous methods. 2) This work is claimed to be the first one on leveraging transformers for tissue microstructure estimation. 3) Extensive comparison with several SOTA methods and the ablation study make the results convincing
- 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
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
This work showed
- 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) The presentation in 2.3 mainly focuses on the mathematical aspect of the RDT module. Adding more explanations about the difference between RDT and standard Unet and the meaning of the notations in terms of imaging parameters will be helpful.
2) More details about the definition of \theta is needed. It is not clear how symmetric q-vectors with respect to the origin and the norm of the q-vector are considered in the definition.
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed work is claimed to be the first one on using transformers for tissue microstructure estimation using diffusion MRI and shows better results than several state-of-the-art methods.
- Number of papers in your stack
3
- 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 #2
- Please describe the contribution of the paper
Deep learning techniques allow prediction of high-quality diffusion microstructural indices from sparsely sampled data. Existing methods are either agnostic to the data geometry in the q-space or limited to leveraging information from only local neighborhoods in the spatial domain. This paper proposes a hybrid graph transformer (HGT) that combines a graph neural network (GNN) and a novel residual dense transformer (RDT), so that the q-space geometric structure and spatial information are fully exploited. Experiments were performed on the HCP dataset for evaluation, where the proposed method has achieved promising results. Overall, this is an interesting paper with proper validation.
- 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.
1) The idea of jointly exploiting the data structure in the q-space and the nonlocal information in the spatial domain is novel. Previous works have not considered the joint use of these different types of information. 2) The selection of competing methods is quite comprehensive. 3) An ablation study was performed to justify the use of residual learning and dense connections.
- 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 threshold of theta appears a bit arbitrary, and it is not clear how sensitive the proposed method is to the threshold.
- 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
The data used in this work is publicly available, but the authors have not indicated whether their method will be made publicly available.
- 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 authors may better clarify what the new contributions in their network design are. Is it the new combination of existing building blocks or are these building blocks also novel?
The authors have not mentioned how the threshold of theta was determined.
Information about standard deviations can be added in Tables 1 and 2.
The authors may consider application of the proposed method to other datasets in the future, especially those with patients.
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I believe that this work proposes a novel method, and its effectiveness has been properly demonstrated. Although there are some minor issues, I recommend acceptance.
- Number of papers in your stack
5
- 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 deep learning approach for estimation of tissue microstructure parameters from sparse diffusion MRI data. They specifically investigate a hybrid graph transformer network that uses both q-space and spatial 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 paper is well written and looks at an important topic. The technical aspects of the method are novel and clearly defined. The experiments are expansive in their evaluation across different approaches.
- 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 experiments focus exclusively on NODDI parameters, which is perhaps something to note earlier. It is not clear how this might generalize to more complex models, e.g. axon diameter or soma density mapping.
- 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
The experiments are described in sufficient detail to be reproduced.
- 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
One thing to consider, when evaluating accuracy of multi-compartment models, you may want to consider the underlying tissue/csf volume fraction. When a voxel is mostly CSF, the ICVF value has little effect of the model fit, so in my opinion, it’s not quite fair to penalize that. You can see in Fig. 2 row 3 that the largest error values are in the ventricles, where ICVF is not very meaningful. One thought is to use a weighted average with weighting by (1 - Fiso) when summarizing error across the image.
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper and underlying work are skillfully executed and a useful addition to the literature.
- 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
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 proposes novel transformer network for the estimation of microstructure parameters from under sampled diffusion MRI data. There is consensus from the reviewers about the technical innovation from this work, and very minor weaknesses.
- 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).
1
Author Feedback
We sincerely thank all the reviewers for their time, efforts, and valuable comments. A point-by-point response is provided in the following.
R1 More details on theta: We map the q-vectors to one hemisphere before computing \theta so that the angle between each pair of q-vectors is less than 90 degrees. We view two q-space sampling points as connected only when they are distributed on the same sphere, i.e., two q-vectors sharing the same norm or b-value. Otherwise, these two q-space sampling points are seen as disconnected.
More explanation about the difference between RDT and the standard UNet: Our RDT differs from the standard UNet in terms of its hybrid architecture and the associated modules. Specifically, our RDT is a hybrid network consisting of a GNN and a Transformer for q-space learning and x-space learning, respectively. In addition, the key modules of RDT are built with a GNN and a Transformer, which are different from the convolution and pooling/unpooling layers used in the standard UNet.
Imaging parameters: We use the data from HCP for training and testing our model. The imaging parameters of HCP will be added to the camera-ready paper.
R2 Novelty: Our HGT is the first hybrid deep learning model for accurate tissue microstructure estimation with joint x-q space information from undersampled DMRI data. It consists of a q-space learning model built with a GNN and an x-space learning model built with Transformer. In addition, we incorporate residual dense connections into the x-space learning model for further performance improvement.
Setting of \theta: We set \theta by performing grid search with different values, i.e., 30, 45, 60, 75, and 90 degrees. The experimental results show that setting \theta to 45 degrees gives the best performance.
Standard deviations in Tables 1 and 2: We will improve Tables 1 and 2 by providing more details.
Application to other datasets: This will be an important future work.
R3 Accurate evaluation of multi-compartment models: We agree with the reviewer’s point that the prior knowledge of multi-compartment models (e.g., the summation of volume fractions equals one) can be considered in the evaluation. Currently, we follow the mainstream evaluation scheme in the field, which is widely adopted, but might not be optimal. We will explore more effective evaluation methods as the reviewer suggested. Furthermore, we will incorporate the prior knowledge of multi-compartment models into the network or postprocessing procedure in the future for further performance improvement.