List of Papers Browse by Subject Areas Author List
Abstract
Structural magnetic resonance imaging characterizes the morphology and anatomical features of the brain and has been widely utilized in the diagnosis of developmental disorders. Given the dynamic nature of developmental disorder progression with age, existing methods for disease detection have incorporated age as either prior knowledge to be integrated or as a confounding factor to be disentangled through supervised learning. However, the excessive focus on age information in these methods restricts their capability to unearth disease-related features, thereby affecting the subsequent disease detection performance. To address this issue, this work introduces a novel weakly supervised learning-based method, namely, the Weakly Supervised Spherical Age Disentanglement Network (WSSADN). WSSADN innovatively combines an attention-based disentangler with the Conditional Generative Adversarial Network (CGAN) to remove normal developmental information from the brain representation of the patient with developmental disorder in a weakly supervised manner. By reducing the focus on age information during the disentanglement process, the effectiveness of the extracted disease-related features is enhanced, thereby increasing the accuracy of downstream disease identification. Moreover, to ensure effective convergence of the disentanglement and age information learning modules, we design a consistency regularization loss to align the age-related features generated by the disentangler and CGAN. We evaluated our method on three different tasks, including the detection of preterm neonates, infants with congenital heart disease, and autism spectrum disorders. The experimental results demonstrate that our method significantly outperforms existing state-of-the-art methods across all tasks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0336_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/0336_supp.pdf
Link to the Code Repository
https://github.com/xuepengcheng1231/WSSADN
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Xue_WSSADN_MICCAI2024,
author = { Xue, Pengcheng and Nie, Dong and Zhu, Meijiao and Yang, Ming and Zhang, Han and Zhang, Daoqiang and Wen, Xuyun},
title = { { WSSADN: A Weakly Supervised Spherical Age-Disentanglement Network for Detecting Developmental Disorders with Structural MRI } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15011},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Authors propose a method based on CGAN to remove the effect of age from the representation of brain scans. They claim that such an embedding should be more discriminative of disorders, such as Autism Spectrum Disorder.
- 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 goal of the paper is clear
- 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.
Methodological choices are not clearly explained
- Please rate the clarity and organization of this paper
Poor
- 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?
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
Not all methodological choices are clearly explained.
- The minimization of the losses brings to a min-max game between the discriminator D (and Q) and the generator G where the generator G should create a representation space from which age can be correctly predicted and that should be the same to the age-related embedding. However, there are no losses (like Mutual Information=0) constraining the disorder-related embedding I to be “disentangled” from the age-related embedding A. And I do not see how the attention module could reach that. Why should I be more discriminative and contain only disorder-related information? Please comment on that.
- The proposed framework resembles to InfoGAN, authors should cite that work.
- Age could also correlate with the disease and integrating age into the analysis, instead than removing it, should also be considered in the comparison. See for instance Dufumier et al. “Contrastive Learning with Continuous Proxy Meta-Data for 3D MRI Classification”. MICCAI, 2021.
- 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 Reject — could be rejected, dependent on rebuttal (3)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please see my comments above.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
It’s still not clear to me how authors can disentangle age-related and disorder-related information using the proposed attention-based method. Furthermore, results seem to point that information is probably not correctly disentangled since in Table 3 diagnostic accuracy slightly improves but the MSE of age prediction drastically increases. My concerns thus remain and I keep my negative score.
Review #2
- Please describe the contribution of the paper
Authors of this paper propose a framework to classify developmental disorders while disentangle and eliminate age-related features from disease-related features. The objective is to design a method that can focus on capturing disease-related patterns instead of an overemphasis on age 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 nicely written and is easy to follow the motivation and proposed framework for the target application.
- Even though accuracy and AUC might not be enough to validate if WSAL module can truly disentangle age and disorder-related features from encoded cortical features, the intuition behind design of WSAL is apparent.
- Authors have included relevant baseline comparisons which provides fair and confident comparison of the proposed framework.
- 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.
- From results it is apparent that the framework can disentangle age and disease related features from cortical features. However, it is not clear if doing so benefits clinical analysis despite the performance improvement. Many neurodegenerative diseases have aging as a potential biomarker, and disentangling that might lead the model to not learn an important biomarker. If there are any strong clinical papers suggesting to not include ageing biomarkers for CHD, authors should include them in introduction.
- It would be insightful if authors can validate the significant brain regions they found with existing medical literature of CHD or through a medical expert. An overlap with existing finding can validate proposed approach with confidence.
- The paper highlights proposed method as a weakly supervised method, however, the corresponding module (WSAL), containing Q network that shares parameters with discriminator D along with the loss objective while using Age values. It is not clear why it is proposed as a weakly supervised method.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
It would be great if authors can address the comments above in weaknesses section which can be useful for readers to better understand the contributions.
- 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?
The proposed method indeed shows a performance improvement and can be useful to many disease detection tasks. Addressing the comments above can be useful to readers.
- 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
Accept — should be accepted, independent of rebuttal (5)
- [Post rebuttal] Please justify your decision
The reviewer appreciates the authors’ responses to provided comments. The responses to Q1 and 3 are acceptable as they clarify the motivation behind age disentanglement in CHD disease and R3 further clarifies an important methodology detail which was confusing otherwise. Authors highlight that they will add additional references as suggested in R2. Considering the novelty of method, application and results, I would consider bumping it up to ‘accept’.
Review #3
- Please describe the contribution of the paper
The method described in the paper reflects the concerns of age-emphasis during inference and training of methods designed to categorise developmental disorders from brain scans. To tackle this, they make use of an attention-based disentanglement module, a conditional GAN and a consistency regularisation to aid in guiding the learning and improve the diagnosis.
- 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 was interesting the assembly of the different modules, in particular, the way to control the influence of age during model inference as well as the architecture choice (spherical Res-Net) which is claimed to be more suitable for the type of data. Evaluations are provided against state of the art approaches and an interesting ablation study to assess the relevance of the attention module and the consistency regularisation term in the loss. In addition to the source code, the authors provide details regarding the implementation, optimisation and hyperparameter selection for those attempting replication of 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 paper would be benefitted from additional details to enhance clarity. There are some definitions missing (e.g. DHCP in page 5, value B in eq. 2, L_{WSAL} does not seem to be explicitly defined elsewere, it just says it is implemented as an L_2-norm, is this between the predicted age from A_i^G and GT age? I would advise to rephrase the first paragraph of page 5 to further clarify this).
It is mentioned that cortical surfaces/features are used as input, but it is unclear how this is obtained, which tools, are used and the type of data present in the three datasets used (i.e. MRI acquisition details), although there is a promise of this to be extended (in the case of in-house data) upon acceptance. Though the use of Spherical Res-Net is interesting, I do wonder if there was a comparison to other more standard architectures.
- Please rate the clarity and organization of this paper
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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Do you have any additional comments regarding the paper’s reproducibility?
The paper has done well in providing an anonimised link to source code as well as implementation details, though dataset preprocessing is not clear, and thus the paper could be enhanced by the inclusion of these details.
- 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
Ther is a clear identification of an issue and the paper addresses the problem in an interesting way. The openness in terms of shared source code and other dataset details is greatly appreciated. There was also an understanding, acknowledgment and evaluation of other approaches to further enrich the confidence of the approach.
- 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?
The paper is very interesting and addresses a potential clinical problem, which is the influence of age in disorder diagnosis. The method proposes an avenue to address this by disentangling age and illness features to enhance diagnosis. The use of consistency term between the output of the attention-based disentanglement and the output of the generator in the conditional GAN was a clever way to
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Accept — should be accepted, independent of rebuttal (5)
- [Post rebuttal] Please justify your decision
The authors provided further clarifications for reviewers’ concerns whilst also showing understanding of the feedback received.
Author Feedback
We thank all the reviewers for the positive comments, such as “nicely written, interesting, useful”. We will release our code to the public at the earliest time after review.
[R1,R3] Q1 Motivation of “disentanglement” strategy: We agree that age can serve as a biomarker for diseases. Our design is not to completely remove age information but only remove a portion of age information correlated to normal developmental components, so that we can enhance the detection of disease-related abnormalities. Specifically, for a patient with a developmental disorder, their brain development consists of normal developmental components (similar to those of HCs) and disease-induced developmental deviations (different from those of HCs). Normal developmental components are highly correlated with age, while disease-induced deviations are more dependent on disease severity. By decoupling these components, our study removes normal developmental information, retaining only disease-related deviations, thereby improving diagnosis accuracy. Additionally, compared to traditional models that confound normal developmental information, focusing solely on developmental deviations enhances understanding of disease mechanisms and extraction of more effective biomarkers. The higher disease detection accuracy and the discovery of more effective biomarker regions in our experimental results both validate the aforementioned hypotheses. We will further clarify the model’s motivation in the final submitted version.
[R1] Q2 The implementation of attention-based disentanglement: Thanks for suggesting mutual information loss, which is an interesting and reasonable idea. We will explore it in future work. In this study, the disentanglement of disease-related and age-related embeddings is achieved by two learning modules based on the attention-based mechanism. Specifically, attention map ψ(z) for disease-related features is optimized through disease diagnosis, while attention map 1 − ψ(z) for age-related representations is optimized through WSAL. The learning and optimization modules of the two channels can assure effective feature disentanglement.
[R3] Q3 Why WSAL is a weakly supervised method: Sorry for the confusion. We use the Q network to predict age from AGi , but not directly for Ai. The learning process of Ai is guided by imprecise AGi . The AGi is imprecise because a) AGi is generated by combining Gaussian noise with age information; b) the Q network makes AGi retain most age information but cannot predict the exact age accurately. Considering the generation guidance of Ai is imprecise, we claim our method as a weakly supervised one.
[R3] Q4 Clinical analysis: The key biomarker regions detected by our method across the three datasets align with existing findings. For example, in the CHD dataset, we found that several regions, including the superior frontal, middle temporal, and right cingulate, are related to CHD neurodevelopmental abnormalities. This is consistent with existing literature (Clouchoux et al., 2013, Cerebral Cortex; Rachael et al., 2014, NeuroImage: Clinical). We will include more references in the final version.
[R1, R4] Q5 Comparison with other methods: Actually, we have tried the method proposed by Dufumier et al. (R1). However, due to the lack of a large scale of cortical morphology data for pretraining, the results are not satisfactory. Therefore, we did not include it and would conduct further exploration in future work. For R4’s suggestion, we replaced Spherical ResNet with CNN and VGG networks and found the ResNet performs best.
[R1, R4] Q6 Writing issues: Due to length constraints, dataset details were omitted in the submitted version but we will include them in the code release. We have also reviewed the paper and corrected issues with missing references(R1), unclear abbreviations, and variable definitions(R4).
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’
The comments from the reviewers are about the motivation behind the method design, clinical analysis, and writing issues. The authors addressed these questions well and indicated that they will correct some errors in the final version. Therefore, I suggest acceptance.
- 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).
The comments from the reviewers are about the motivation behind the method design, clinical analysis, and writing issues. The authors addressed these questions well and indicated that they will correct some errors in the final version. Therefore, I suggest acceptance.
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’
Despite some issues regarding the clarity, this paper introduces interesting ideas that are of interest to MICCAI community. The rebuttal addresses most of the issues, and I hope the final paper is more clear.
- 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).
Despite some issues regarding the clarity, this paper introduces interesting ideas that are of interest to MICCAI community. The rebuttal addresses most of the issues, and I hope the final paper is more clear.