List of Papers Browse by Subject Areas Author List
Abstract
Cortical parcellation delineates the cerebral cortex into distinct regions based on anatomical and/or functional criteria, a process crucial for neuroscientific research and clinical applications. Conventional methods for cortical parcellation involve spherical mapping and complex feature computation, which are time consuming and prone to error. Recent geometric learning approaches offer some improvements but may still depend on spherical mapping and could be sensitive to mesh variations. In this work, we present Cortex-Diffusion, a fully automatic framework for cortical parcellation on native cortical surfaces without spherical mapping or morphological feature extraction. Leveraging the DiffusionNet as its backbone, Cortex-Diffusion integrates a newly designed module for full-band spectral-accelerated spatial diffusion learning to adaptively aggregate information across highly convoluted meshes, allowing high-resolution geometric representation and accurate vertex-wise delineation. Using only raw 3D vertex coordinates, the model is compact, with merely 0.49 MB of learnable parameters. Extensive experiments on adult and infant datasets demonstrates that Cortex-Diffusion achieves superior accuracy and robustness in cortical parcellation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2548_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2548_supp.pdf
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Zhu_Efficient_MICCAI2024,
author = { Zhu, Yuanzhuo and Lian, Chunfeng and Li, Xianjun and Wang, Fan and Ma, Jianhua},
title = { { Efficient Cortical Surface Parcellation via Full-Band Diffusion Learning at Individual Space } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15002},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a fully automatic framework for cortical parcellation on native cortical surfaces without spherical termed as Cortex-Diffusion. This method, unlike related works, tries to obtain cortical parcellation without spherical mapping and morphological feature quantification. To achieve this, the framework incorporates DiffusionNet and a newly designed module, capturing high-resolution, detailed geometric information. Experiments conducted on validation sets validate the effectiveness of their contributions.
- 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）Innovation: Introducing a full-band spectral-accelerated spatial diffusion learning module represents a significant advancement over traditional methods that require complex pre-processing and spherical mapping. 2）Efficiency: By avoiding spherical mapping and direct use of raw 3D coordinates, the model is not only efficient in computational resources (0.49 MB of learnable parameters) but also robust in handling high-resolution data efficiently. 3）Accuracy and Robustness: Extensive testing on both adult and infant datasets demonstrates superior performance in cortical parcellation compared to existing methods, underlining both the method’s accuracy and robustness.
- 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) While the paper claims efficiency, the underlying computational complexity and the need for specific hardware (e.g., capable GPUs) for optimal performance are not thoroughly discussed. 2) The paper focuses on cortical parcellation but does not discuss the applicability of the method to other similar tasks in medical imaging, which could limit its broader impact. 3) Some aspects of the full-band spectral-accelerated module are briefly covered. A more detailed exposition on the adaptability of the module across different mesh resolutions and variations could enhance the paper’s comprehensiveness.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
1) It would be beneficial if the authors could provide more details about the model’s performance on standard computational resources, not just specialized hardware, to better understand its applicability in typical clinical settings. 2) The paper would benefit from a deeper dive into how the full-band spectral-accelerated spatial diffusion is computed, especially how it handles the transitions between different frequencies and how it affects the accuracy and efficiency of the parcellation. 3) Releasing code after acceptance can greatly enhance the impact and reproducibility of the paper. 4) minor suggestion: I am concerned about the quality of surface reconstruction of dHCP using Freesurfer.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making
Reject — should be rejected, independent of rebuttal (2)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a significant advancement in the field of cortical surface parcellation with its novel approach. However, to strengthen the paper and justify its findings more robustly, addressing the detailed comments provided would be crucial.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
I am still concerned about the reproducibility.
Review #2
- Please describe the contribution of the paper
The paper introduces a novel Cortex-Diffusion framework for cortical surface parcellation leveraging DiffusionNet and a newly designed module for full-band spectral-accelerated spatial diffusion learning. This method aims to improve the accuracy and efficiency of cortical parcellation without relying on spherical mapping or morphological feature extraction, using only the raw 3D vertex coordinates as input. The authors report that their model is compact, with a mere 0.49 MB of learnable parameters, and demonstrate its performance through extensive testing on adult MindBoggle and infant dHCP datasets.
- 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 integration of a full-band spectral-accelerated spatial diffusion learning module represents a significant technical advancement. This allows for adaptive information aggregation across highly convoluted meshes, potentially capturing high-resolution, detailed geometric information critical for accurate vertex-wise delineation of cortical regions.
– The model’s efficiency is notable, requiring only 0.49 MB of learnable parameters and obviating the need for time-consuming processes such as spherical mapping and morphological feature computation.
– The paper presents convincing evidence of the model’s robustness and accuracy across various datasets and conditions, especially highlighting its performance in the context of discretization-agnostic properties and its robustness to changes in mesh resolutions and connectivity patterns.
- 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 paper introduces the Cortex-Diffusion model for cortical surface parcellation, tested on two datasets with an ablation study to demonstrate its architectural contributions. This approach shows the model’s effectiveness and potential applicability across different contexts. Yet, it leaves open questions on the baseline methods chosen for evaluation.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
– The design focusing solely on 3D vertex coordinates might miss capturing intricate cortical geometries that are more effectively achievable with advanced features.
– The introduction of FB-SASD offers technical innovation but the discussion needs more depth on its computational demands versus accuracy improvements, particularly concerning eigen-decomposition.
– Robustness against mesh variations are thoroughly tested. Does the method work under extreme deteriorations or topological errors like self-intersections or holes. Does the method rely on pre-processing like FreeSurfer for topology correction. If so, how efficient would the method be compared to the recent advancement in deformation based cortical surface reconstruction and parcellation. Eg. Vox2Cortex or Bongratz, Fabian, et al. “V2C-Long: Longitudinal Cortex Reconstruction with Spatiotemporal Correspondence.”
– The method’s handling of high-frequency information through nonlinear mapping is innovative but needs to be empirically justified, especially regarding its specific contribution to improving parcellation accuracy.
– The paper does not adequately compare its proposed method against a broad spectrum of existing approaches that utilize spectral, spherical, and unsupervised learning techniques for cortical surface parcellation. Specifically, it overlooks direct comparisons with notable spectral methods like Gopinath et al.’s Spectral GCN (2018), fails to assess advancements in spherical convolutional neural networks as seen in Parvathaneni et al. (MICCAI 2019), and does not consider the implications of recent unsupervised learning and denoising approaches in cortical surface analysis, such as those by Cheng et al. (NeuroImage, 2020) and SI Young et al. (PAMI, 2023). Incorporating these comparisons could have enriched the discussion on the proposed method’s position within the current state-of-the-art, offering a clearer perspective on its novelty, efficiency, and potential advantages over existing techniques. The choice to not compare against point cloud processing methods like PointNet or PointNet++ for baseline validation misses an opportunity to contextualize the model’s performance against established geometric deep learning approaches.
– Figures 1 and 3 are crucial for understanding the methodology and appreciating the qualitative results; however, they are not easily understandable if these figures are too small.
– Typo: QEM algorith
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation acknowledges the paper’s contribution to advancing cortical surface parcellation techniques while highlighting areas where further work could strengthen its claims and broaden its applicability. The balance of innovative methodology against the outlined limitations and opportunities for deeper comparative analysis justifies a “weak accept” rating.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
Introducing a full-band spectral-accelerated spatial diffusion learning model for cortical parcellation to the MICCAI community is interesting. The paper is written well and lacks comparison to some spectral methods. However, the authors have addressed the rebuttal well.
Review #3
- Please describe the contribution of the paper
In this work, the authors present an application (Cortex-Diffusion)) of the method (DiffusionNet) for the cortical parcellation on native cortical surfaces without spherical mapping or morphological feature extraction. With DiffusionNet as the backbone architecture, the proposed method uses a module for full-band spectral-accelerated spatial diffusion learning to adaptively aggregate information across highly convoluted meshes. The method is tested on adult and infant datasets: Mindboggle and dHCP.
- 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.
- It is a novel application of an established method in the graphics community to parcellate the brain in the native space itself rather than doing transformations to some standard template.
- The results are comparable to the state of the art methods. They do not outperform the other methods by any large margin but that is okay. The method solves an important problem in native subject space.
- 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 paper can tighten-up some of the language as follows: Abstract: “involve spherical mapping and complex feature computation, which are time-consuming and prone to error.” What particular errors do the authors refer here?: Please provide some more details. In the introduction, the claims (1) and (2) are not backed up: “1) The involvement of multiple mapping and registration phases may lead to the introduction of unnecessary errors; 2) The process of spherical mapping, along with the computation of cortical features, is notably time-intensive, typically requiring more than an hour per subject on a standard PC.” (1) The phrase “unnecessary errors” is vague and needs more context. What specific errors are the authors talking about? (2) Though this might be true, it not entirely correct. Just the spherical mapping process does not usually take more than an hour can speed depending on the number of threads. The topologically corrected surface inflated to sphere and then spherical registration doesn’t take much time, it is the surface-based parcellation in the next step which is the most time consuming.
- The evaluation is in terms of Dice. I am not sure how Dice can be used to compare cortical surfaces? Were the parcellations projected to the volumetric space?
- The Dice improvement is marginal and also this metric might be incorrect for surface-parcellations?
Minor: Typing error on page 6: “QEM algorith” should be “QEM algorithm”.
- Please rate the clarity and organization of this paper
Excellent
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
The authors did not comment if they will open-source 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. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
As mentioned, the authors could tighten-up some loose ends with some description on the claims made as pointed out in the weaknesses section.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is a good paper but authors could provide more details on the evaluation done and how else the method be evaluated except based on Dice score.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Reject — should be rejected, independent of rebuttal (2)
- [Post rebuttal] Please justify your decision
After reading the reviews from the other reviewers as well and the author feedback, I still like the author’s approach but I reserve doubts about the lack of substantial improvements of the results on the datasets as the method is more of an application of an existing approach applied to this new setting of cortical parcellation. The other reservation is that since the method is dependent on only the vertex coordinates, what would happen in the areas where the underlying brain tissue was not segmented by FreeSurfer and hence the associated mesh might have missing regions (aka vertex coordinates) in that brain region. The method was applied to a healthy population and doubt remains on a diseased population which is more challenging it segment and parcellate. This is something the authors did not discuss.
Author Feedback
We thank all reviewers for their constructive comments and appreciate their praise for our paper’s contribution. R1: the paper makes “contribution to advancing cortical surface parcellation techniques”. R3: the paper presents “a novel application of an established method in the graphics community to parcellate the brain in the native space”. R4: “The paper presents a significant advancement in the field”. We addressed the major concerns below:
*[R1]-Performance with advanced morphological features We’d like to emphasize that the goal of our study is to develop a lightweight and stable method for cortical parcellation with as-simple-as-possible precomputations, due to this reason only vertex coordinates were used as the input. Notably, if morphological features are involved, our method could further improve the performance. In contrast, Spherical U-Net and SPHARM-Net using morphological features performed worse than ours using solely coordinates.
*[R1&R4]-Computational demands & complexity The computation cost is related to the number of eigenvectors (i.e. k) and the resolution of cortical surface. Specifically, it takes 44 seconds to decompose a cortical surface with 130,921 vertices when k=200 (the default setting in our experiments), which is acceptable. More importantly, we should indicate that our method is also very efficient on standard computational resources. It takes only ~4.5 seconds to parcellate a high-resolution surface with 141,802 vertices on a standard PC with an Intel Core i7-11700 CPU.
*[R1]-Effectiveness under extreme conditions Our experiment was conducted on publicly released datasets, under the assumption of no topological errors. Considering that our method processes directly original surfaces, it by nature works also under topological errors, which is an advantage compared with spherical networks. We pursue to check the effectiveness under such conditions in the future.
*[R1&R4]-Empirical justification for nonlinear mapping of high-frequency information & efficiency of full-band spectral accelerated spatial diffusion At the end of section 3.2, we conducted an ablation experiment, in which we compared the performance of two kinds of spatial diffusion mechanism across different situation, demonstrating our design’s superiority.
*[R1]-Comparison with existing methods We merely aimed at proposing a backbone that can be directly performed on cortical surface represented by mesh, foregoing the need for spherical mapping. Therefore, we mainly compared with spherical mapping-based methods, i.e. Spherical U-Net and SPHARM-Net, where the latter is a more advanced spherical convolution network than Parvathaneni et al., and mesh deep learning method, i.e. SubdivNet. Although supervision manner is not what we are interested in, we can combine our model with unsupervised learning mechanism in future work. And we will consider adding a comparison with Spectral GCN. Directly comparing with point cloud methods is unfair because they typically search for nearest neighbors in Euclidean space for convolution and are therefore not suitable for highly convoluted cortical surfaces, while we can utilize the predefined connectivity of mesh.
*[R3]-Using Dice as the evaluation metric We calculate the Dice score according to its standard formula in vertex-level space rather than volumetric space. Notably, Dice is a commonly used metric by almost all methods in this field.
*[R3]-More detailed description to support our claim Following the suggestion, we’ll update the manuscript to describe our claims more clearly.
*[R4]-Surface reconstruction of dHCP We directly used the surface files released by dHCP project, and we’ll update relevant description in the manuscript.
*[R1, R3, R4]-Reproducibility We’ll release our code, and it is currently not available due to anonymity. The datasets we used are all publicly available.
Meta-Review
Meta-review #1
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
- 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).
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
- 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).
N/A