List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Xiaoyang Chen, Junjie Zhao, Siyuan Liu, Sahar Ahmad, Pew-Thian Yap
Abstract
The baby brain undergoes rapid changes in volume, shape, and structural organization during the first postnatal year. Accurate cortical surface reconstruction is prerequisite to characterizing dynamic variations in cortical morphometry during early brain development. Existing surface reconstruction methods are typically tailored for adult brain MRI and are inaccurate in reconstructing cortical surfaces from infant MRI, owing to the poor tissue contrast, partial volume effect, and rapid changes in cortical folding patterns due to the emergence of secondary and tertiary cortical folds in the first year of life. Here, we propose an infant-centric cortical surface reconstruction method that accommodates dynamic variations in early developing brain. Our method, SurfFlow, sequentially deforms an initial template mesh to target cortical surfaces with 3 seamlessly connected deformation blocks. It can rapidly reconstruct a high-resolution cortical surface mesh with 460k vertices in about one second. Performance evaluation based on a dataset of infants from 0 to 12 months indicates that SurfFlow reduces geometric errors and improves mesh regularity substantially over state-of-the-art deep learning approaches.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43993-3_37
SharedIt: https://rdcu.be/dnwNC
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel method for reconstructing high-resolution cortical surface, which result in reducing geometric errors and improving mesh regularity substantially.
- 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.
- This paper proposed a multi-block model to predict the flow field.
- This paper is well-organized and written.
- By comparing to CorticalFlow, the proposed method achieved better 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.
- It might be a great challenge to extract cortical gyral and sulcal patterns in infant brain image. However, this paper didn’t show the motivation why infant is difficult and important.
- Figure 2 is hard to interpret. It would be nice to show zoom-in views near the landmarks.
- 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 description of the SurfFlow implementation is too sketchy to be reproduceable, unless authors (really) publish their code.
- 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/2023/en/REVIEWER-GUIDELINES.html
The following issues should be considered:
- In the Sec. 2.4, the training process of 3 blocks are not clear. are these three blocks are trained separately?
- The evaluation is too brief to clarify how and why this evaluation is need.
- 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?
Important details are missing to understand the work presented here. Authors do not provide sufficient quantitative evidence to prove their claims.
- 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
3
- [Post rebuttal] Please justify your decision
I hold my original opinion due to the lack of innovation, low quality graph and insufficient evaluation.
Review #2
- Please describe the contribution of the paper
The authors proposed an cortical surface reconstruction with direct application for infant brain dataset. They proposed method introduces an efficient dual-modal flow-based cortical surface reconstruction technique that generates high-resolution and high-quality mesh representations of complex cortical surfaces. A new loss function is introduced to regularize mesh edge lengths, resulting in significant improvements in mesh quality. This model is the first to directly reconstruct surfaces from challenging infant brain MRI, outperforming current state-of-the-art methods by a considerable margin across various surface evaluation metrics.
- 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 paper is clear, well written, and understandable. 2- The results section is very good, quite extensive experiments for a conference paper and showing different aspects of the results. 3- The contributions are well articulated and shown in the paper. The paper delivers on promises in the introduction.
- 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- Figure 1 is not well done, the figure in the supplementary material would be better if merhged. 2- The authors do not handle space well especially for figure 2 and 3, some results could have gone to the supplementary material more andwhile keeping the most important results in the paper. 3- Related to the previous, this resulted in a brief discussion that could have been extended.
- 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 paper seems to be well reproducible.
- 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/2023/en/REVIEWER-GUIDELINES.html
1- The authors should have improved alot the figure 1. Especially including details. I strongly believe that figure 1 and the model figure in the supplementary material could be merged in one figure in the same space as figure 1 under some good space management.
2- the discussion should have been better articulated.
- 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 authors put a set of significant contributions in the paper. The writing and clarification of the model are good. The loss function idea is very good.
- 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 #4
- Please describe the contribution of the paper
The paper proposes a geometric deep learning model to reconstruct cortical surfaces from infant brain MRI, which includes a new loss function to regularize mesh edges. On the BCP dataset, the method produces large improvement over CorticalFlow. Significant ablation studies were performed.
- 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.
Paper is clearly written, well-motivated and easy to follow. Good clinical application. Parts of the model are well ablated in table 2. Good quantitative results.
- 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 new proposed loss is barely motivated.
- The method only uses a single baseline (CorticalFlow) - other approaches outlined in the introduction such as FreeSurfer, DeepCSR, Voxel2Mesh, PialNN should be used.
- The method is applied only on a single dataset, thus it is unclear how well the proposed method will generalize.
I look forward to the authors’ responses. Please put any changes or additions you would make (new results, new baselines) in the rebuttal.
- 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. Would be much strengthened if the authors released the code to the reviewers.
- 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/2023/en/REVIEWER-GUIDELINES.html
- The authors have much space remaining in the main paper and could move elements of the supplementary materials into the main paper
- Tables S1, S2 in the supplementaries are not referenced in the main paper.
- 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
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please see ‘main weaknesses’ section.
- Reviewer confidence
Somewhat confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [Post rebuttal] Please justify your decision
Thank you for the response.
‘In addition, we will publish the code on GitHub.’ I recommend in future papers the authors present an anonymized version of the code.
‘Due to space limitation, we only compared CorticalFlow, the current SOTA, with our method…[extra results]’ I appreciate the single set of extra results. I still believe another 1-2 comparisons would be useful. I am pleased the authors outperformed the SOTA CorticalFlow – and recommend the authors add this is why they used CorticalFlow as a baseline. This paper is well below the page limit, so the authors have space to add extra results.
‘The naive edge length loss cannot..’ I think the proposed loss requires much more motivation - what the authors posted is still not sufficient. I agree with the authors that the loss is useful, however/
’- Single-dataset experiments make the generalizability of method unclear (R4)’ (my comment) I was requesting performance numbers on these new datasets.
I am raising my score to accept as I think this paper tackles an important problem that is not well-addressed by current methods. However, as most of my comments have not been addressed (only results on a single dataset, lack of motivation, odd presentation) I am not going above a borderline/weak accept. I also expect the authors to release the code as promised, if the paper is accepted.
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.
Several reviewer comments - requires author rebuttal.
- reviewer 2 has some moderate concerns, please make sure you address them.
- some explanation on challenges for infant mri should be added.
- There are some concerns about organization and presentation of results by reviewer 2 and 3.
Author Feedback
Why is infant cortical surface reconstruction (CSR) difficult and important (MetaRev, R1) Compared with adult MRI, CSR using infant brain MRI is significantly more challenging due to the lower contrast and larger temporal variation due to rapid brain development (Sec 2.2). Infant MRI can also suffer from lower signal-to-noise ratio and more pronounced motion artifacts (doi.org/10.1016/j.neuroimage.2018.04.044). The reconstructed surfaces facilitate brain morphology (e.g. cortical thickness and sulcal depth), surface-based analysis, and visualization of brain function and anatomy (doi.org/10.1136/jamia.2001.0080443).
Implementation details for reproduction (R1, R4) To improve the clarity, we will include a paragraph to describe in detail how the flow ODE is solved numerically by integrating the stationary flow field to obtain the deformation field via the fourth-order Runge-Kutta method. In addition, we will publish the code on GitHub.
Comparison with other methods (R4) Due to space limitation, we only compared CorticalFlow, the current SOTA, with our method. Further experiments with DeepCSR indicate that the Chamfer Distance (CD) value is 4.69 ± 2.10 on the left pial surface; 3.85 ± 2.00 on the right pial surface; 0.74 ± 0.50 on the left white surface; and 0.50 ± 0.90 on the right white surface. The pial surface error is notably higher due to numerous topological artifacts, even after topology correction. Quantitative evaluation shows that the SurfFlow performs the best among the 3 methods.
Description of evaluation and discussion of results too brief (R1, R2, R4) We have extended our discussion in the results section. We find that SurfFlow consistently performs the best in all metrics. The lower CD (<0.5 mm vs >1 mm) indicates a smaller error compared with the other two methods. SurfFlow is also more robust with lower std. PELL encourages mesh edge lengths to lie within the desired range. With PELL, predictions by SurfFlow are more uniform and accurate. By contrast, CorticalFlow is unable to reconstruct some surface areas (e.g., sulci), mainly due to its limitation in ensuring mesh uniformity (Fig 4). The Normal Consistency metric also suggests that our method is better in terms of surface smoothness.
Network structure should be merged into Fig 1 (R2) As suggested, we tightened the vertical space of two figures, shrunk the Unet figure inside supplementary material and merged them to form a complete network structure figure. Now the network structure figure in the main paper includes all details.
Lack of zoom-in details of Fig 2 (R1) We added a figure to show the lateral view with more detail of the reconstructed white surface of one subject, comparing SurfFlow with CorticalFlow, DeepCSR, and GT.
Move some from Supp to the manuscript (R2, R4) We merged pial and white numerical evaluation results and added the comparison to DeepCSR in Table 1. We also combined Table 2 in supplementary material with Table 2 in the main paper. Error map figure for the pial surface is moved to the main paper.
Are blocks in Sec 2.4 trained separately? (R1) No. We stated in Sec 2.4 that these three blocks are trained sequentially by training one block at a time while freezing the weights of the other U-nets.
Motivation of new proposed loss (R4) The naive edge length loss cannot ensure evenly distributed vertices, while PELL constrains the edge length to a desired range, enabling mesh uniformity (Sec 2.3). We show PELL improves performance via an ablation study (Table 2).
Single-dataset experiments make the generalizability of method unclear (R4) Generalizability can be tested based on larger datasets for example from the HBCD study and NIH Pediatric MRI Data Repository. Generalizability can also be improved via data augmentation in a manner similar to SynthSeg (doi.org/10.1016/j.media.2023.102789).
Lacking citations of Tables S1, S2 in manuscript (R4) We will merge them into the main paper and discuss the result.
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.
Comparison with other methods was major reviewer concern. The author in their rebuttal have addressed this issue. Some other concerns about low quality figures, implememtation details are relatively minor in my opinion, and authors have promised to address those in the final version.
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.
This paper proposes a flow-based method for cortical surface reconstruction from infant MRI data, which is known to be a challenging problem. The main idea is to deform an initial template with a diffeomorphic displacement field that is parameterized as the end point of a flow ODE. The main strength of the paper is that the proposed approach clearly outperforms the chosen baseline method CorticalFlow on the Baby Connectome Project data. However, this is also leads to two major weaknesses of the work: (1) no other baselines were considered and (2) the approach is only evaluated on a single database, which raises question regarding the generalizability of the method.
The authors justify the lack of inclusion of additional baseline methods with space restrictions, but I think that’s a rather weak argument here as it would have only taken a couple of additional rows in the results table to achieve this. Moreover, it is claimed that CorticalFlow represents the SOTA. However, CorticalFlow++ ([10] - a MICCAI 2022 paper) significantly outperformed CorticalFlow in [10]. It is therefore unclear if the proposed method is really tested against the SOTA. Moreover, I am struggling a bit to see much technical novelty in the paper. The whole setup appears to be highly similar to that of CorticalFlow++ with two notable exceptions: (1) The proposed method can utilize two MRI sequences and (2) the authors propose a new loss function. However, and as pointed out by R#4 the proposed loss function is barely motivated and the integration of a second modality into the CoriticalFlow/CorticalFlow++ setup seems to be a rather trivial extension. For me those weaknesses unfortunately outweigh the presented performance gains in comparison to CorticalFlow.
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.
One of the reviewer’s concerns were not adequately addressed. For another reviewer, the low rating was either maintained, or borderline. The method was not adequately motivated and only used a single baseline method (CorticalFlow). The generalizability of the method was in question as it was only tested on a single dataset. Important details about the method were missing although the authors responded that they will add those later. The authors should pay close attention to the reviewers comments and update the paper with all the important changes suggested in the rebuttal for the final submission.