List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Huabing Liu, Dong Nie, Dinggang Shen, Jinda Wang, Zhenyu Tang
Abstract
Many deep learning (DL) based methods for brain tumor segmentation have been proposed. Most of them put emphasis on elaborating deep network’s internal structure to enhance the capacity of learning tumor-related features, while other valuable related information, such as normal brain appearance, is often ignored. Inspired by the fact that radiologists are often trained to compare with normal tissues when identifying tumor regions, in this paper, we propose a novel brain tumor segmentation framework by adopting normal brain images as reference to compare with tumor brain images in the learned feature space. In this way, tumor-related features can be highlighted and enhanced for accurate tumor segmentation. Considering that the routine tumor brain images are multimodal while the normal brain images are often monomodal, a new contrastive learning based feature comparison module is proposed to solve incomparable issue between features learned from multimodal and monomodal images. In the experiments, both in-house and public (BraTS2019) multimodal tumor brain image datasets are used to evaluate our proposed framework, demonstrating better performance compared to the state-of-the-art methods in terms of Dice score, sensitivity, and Hausdorff distance. Code: https://github.com/hbliu98/CLFC-Brain-Tumor-Segmentation
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16443-9_12
SharedIt: https://rdcu.be/cVRyq
Link to the code repository
https://github.com/hbliu98/CLFC-Brain-Tumor-Segmentation
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents the brain tumor segmentation framework by adopting normal brain images as a reference to compare with tumor brain images in the learned feature space. The main contribution of the proposed method is to use monomodal normal brain images as a reference to improve the segmentation performance and the contrastive learning-based feature comparison (CLFC) module that is designed to solve the incomparable issue between features learned from multimodal tumor brain images and monomodal normal brain images. The proposed method has been validated on two datasets which include the in-house dataset and the publicly available Multimodal Brain Tumor Segmentation (BraTS2019) dataset.
- 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 proposed method take advantage of both multimodal and monomodal data to increase the performance of tumor segmentation.
-
CLFC module, seems to be simple, yet effective. It improves the efficiency of segmentation backbone.
-
The experimental results show that the proposed method improves segmentation over the method which does not use the contrastive learning.
-
- 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 only comparison technique is nnUNet. The Brats2019 data labels were not compared for each, so it is difficult to compare it with other techniques of the leaderboard.
-
The results show that the proposed method performs well on binary segmentation task, but how does it perform for multi-class segmentation? This is what I am more concerned about because it seems difficult for normal images to help distinguish the types of tumors. Please provide some more details about it.
-
Interestingly, the baseline 1 has already outperformed nnUNet even though it consists of a general Unet structure. It is necessary to explain how such a result can be obtained.
-
Compared to the baseline methods, the improvement of the proposed method is not large.
-
- 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
Authors checked “Yes” for most questions on the reproducibility of the paper
- 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
-
It would be interesting if the proposed method is compared with other contrastive learning methods such as MoCo (https://arxiv.org/pdf/1911.05722.pdf ) and SimCLR (https://arxiv.org/pdf/2002.05709.pdf ) instead of SimSiam network.
-
Usually, contrastive learning methods require a large number of negative samples to avoid the mode collapse, the author used SimSiam which use only positive sample. Then the proposed method should be compare with other the contrastive learning methods which only uses positive samples such as BOYL(https://arxiv.org/pdf/2006.07733.pdf ).
-
The Normal Appearance Network performance depends on the reconstructed monomodal normal brain images from IntroVAE. It would be better to show qualitative and quantitative analysis of the generated samples.
-
Please also explain if the segmentation backbone and Normal Appearance Network were trained from scratch or these are pre-trained models.
-
The author should also include the hyperparameter sensitivities for the proposed method in the training details
-
- 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?
It is an interesting method for tumor segmentation that utilizes both multimodal and monomodal data to improve the performance, but there are lacks experiments to demonstrate its effectiveness against the baselines and the existing methods based on the contrastive learning.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
4
- [Post rebuttal] Please justify your decision
I hold my opinion because additional experiments need to be included to prove the effectiveness of each element. In the case of BratS data, multi-class classification results also need to be compared.
Review #2
- Please describe the contribution of the paper
In this paper, the authors tackle the task of brain tumor segmentation from multimodal Magnetic Resonance Imaging (MRI) using Convolutional Neural Networks (CNNs). The main contribution is in utilizing MRI images of normal brains to help the network in contrasting the tumor to the normal region. The normal images are generated at runtime by reconstruction from IntroVAE. To further enhance learning, the authors propose that the features of normal regions are aligned between the features of the brain with tumors and features of normal brains. Finally, an attention map is generated to enhance tumor features. The method is evaluated in the publicly available Brain Tumor Segmentation (BraTS) Challenge 2019 dataset and in an in-house dataset. Results of the proposed method improve over the baselines.
- 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 writing of the manuscript is generally good and clear.
- To the Reviewer’s knowledge, the idea of using normal brain images to help brain tumor segmentation with CNNs is novel. This is further enhanced by the feature alignment module for regions of the normal brain. This idea is inspired by how radiologists learn to identify brain tumor, i.e., by knowing well the normal brain structure. Incorporating this concept is interesting.
- The ablation studies show the effectiveness of the proposed method over the baselines.
- 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 proposed method has a lot of components such as 1) segmentation network, 2) normal tissue network, 3) IntroVae, and 4) feature alignment module. This makes it harder to implement the method, especially because IntroVAE needs training just for itself.
- The authors claim to use Contrastive Learning to align the features, but the employed method is SimSiam, which is a non-contrastive learning method [1] (i.e., it only uses positive pairs). This requires adapting the text and title.
- The proposed method only deals with whole tumor segmentation task as a binary task. But, in BraTS, it is well-known that a brain tumor can be divided into 3 major regions: enhancing tumor (ET), Tumor Core (TC), and Whole Tumor (WT). The proposed method would be stronger if it was shown to address these classes.
- The results are obtained in a five-fold cross-validation way. Therefore, we can assume that they are computed from the held-out fold in each run. This can be susceptible to overfitting and overly optimistic results. In other words, a separate test set would be important.
- 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
The proposed method contains many components, which makes it harder to reproduce from the implementation side. Still, in the Reproducibility Form, the authors promise to make the code available, which may help. Additionally, no details are provided about training IntroVAE. while this is not the focus of the work, it is a key component, so, details would be needed. The authors also use an in-house dataset for part of the results, therefore, it is not possible to reproduce these results.
In summary, this Reviewer considers that there are some risks, and reproducibility potential is just moderate.
- 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
General
This paper focuses on the task of brain tumor segmentation in MRI, and the main idea and contribution is the use of normal brains to contrast with the brains with tumors. To that end, the authors learn an IntroVAE from normal T1 scans and generate a normal brain at runtime. During training, the authors propose a loss that aligns the features of normal regions between normal and tumor scans, such that features of the tumor are enhanced by an attention map. The paper is well-written and the method is novel. However, there are some concerns such as dealing with the “Whole Tumor” class only or the evaluation procedure.
Comments/questions to the authors (not in order of importance)
1) The proposed method has a lot of components, such as the segmentation network, but also the IntroVAE for generating normal brain images. a) There are no details about the training of IntroVAE. It is also unclear how much the segmentation network is affected by this model. Could the authors add supplementary material with the recipe to train the IntroVAE, please? b) Related to the previous question, could the authors provide some images with the reconstructed images, especially in the tumor region, please? c) Could the authors release the source code, please? It would help the readers in implementing this approach.
2) The evaluation procedures raise some concerns. a) The authors use five-fold cross-validation. While this is a valid approach to searching hyper-parameters, it is fair to assume that the reported metrics are the average of the held-out folds, used for validation during each training. Therefore, it can suffer from over-fitting and the results may be overly optimistic. In other words, it would be desirable to have an independent test set. b) Related to the previous question. BraTS makes a validation set that can only be evaluated online with hidden ground truth. Could the authors submit their method and the baseline and provide such results, please? This wouls allow to compare with the state of the art. c) The authors employ Wilcoxon signed-rank test and report p-values to compare across methods. However, the significance level must be mentioned, otherwise, the reader does not know the condition to verify the hypothesis. d) No comparison with SotA is provided, other than comparisons with the baselines. Although the authors only tackle the Whole Tumor class, comparisons could still be made in these classes.
3) The authors only report the metrics about the binary whole tumor segmentation. However, in BraTS, the tasks also include Core Tumor and Enhancing Tumor. a) The proposed paper would be stronger if the authors extend the method to the other classes. Why did the authors choose to tackle Whole Tumor? Could this be made clear in the manuscript, please?
Further comments (suggestions/extra comments on future work) - NOT intended to be addressed during rebuttal
1) The proposed method is interesting and new. Here are some suggestions for an extension of the work that the Reviewer considers would make it a stronger work. As an extension of the work, it would be good to train the network with Core Tumor and Enhancing Tumor and report these metrics, too. Also, a comparison with the SotA in BraTS 2019 is necessary to show the benefits of the proposed work. Finally, the authors propose to generate a normal brain from the patient’s acquisition, which is interesting. However, it is not clear how much the method is sensitive to this. A simple experiment would be to replace the generated brains with an Atlas of T1 sequence, registered to the patient’s scans, and check the differences.
References
[1] Tian, Yuandong, Xinlei Chen, and Surya Ganguli. “Understanding self-supervised learning dynamics without contrastive pairs.” International Conference on Machine Learning. PMLR, 2021.
- 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?
The work proposed in the paper is interesting and new. However, there are some weaknesses, such as the evaluation procedure, the setting of just segmenting Whole Tumor that is different from the setting in BraTS and the brain tumor segmentation community, and the imprecise use of the Contrastive Learning concept, when what was used was a non-contrastive learning method (SimSiam). So, despite the interesting work, the Reviewer believes that some changes would be necessary.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
2
- 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
5
- [Post rebuttal] Please justify your decision
Thank you for submitting the rebuttal and addressing the comments.
1) The authors address some of the initial concerns regarding using five-fold cross-validation and potential overfitting. The Reviewer is convinced that good practices were followed.
2) In the opinion of the reviewer, the reason to employ binary segmentation is not convincing. That can be the case of the in-house dataset, but BRATS allows for a multi-class setting. Therefore, a different setting could be done in each dataset, since the analysis is done separately.
3) Thank you for the details about introVAE. The Reviewer agrees that the results seem stable. It would be good to make such information available upon publication, e.g., in suplementary material or in the code repository.
4) SimSiam does implicit contrastive learning. Nonetheless, the community converged to differentiate the types of models and that should be followed.
The authors address some concerns, while others are not so satisfactory. However, the Reviewer is improving the rate to “weak accept” in recognition to the explanations and the merits of the paper.
Review #3
- Please describe the contribution of the paper
This paper proposed a novel deep learning segmentation model to deal with brain tumor segmentation in multimodal MR images. The novelty of the proposed model is to integrate the appearance of healthy brain MRI in order to localize the anomalies that is further used to strengthen the segmentation accuracy. In specific, capturing the appearance of normal-looking brains does not necessarily require all four MR sequences as the model relies only on T1 MR sequence of the subjects. The normal-looking T1 images are synthesized from the original pathological T1 sequences through an IntroVAE model. To tackle the inconsistency between the multimodal image data and the mono-modal normal appearance data, a contrastive learning-based module was designed as well. The model was developed and tested on 2D orthogonal views of the BraTS19 dataset and compared against one standard segmentation model in which the superiority of the model is highlighted in terms of segmentation accuracy.
- 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 well-developed the following points: 1)a novel method to generate normal appearance MR slice from the corresponding tumoral slice. 2)they employed an advanced model to efficiently capture the information from so called paired images of healthy-unhealthy slices. 3)Extensive experiments were conducted to evaluate the efficacy of the model. In general, the studied problem is of great interest for the MICCAI community, the proposed method sounds both theoretically and experimentally, and the reported results on comprehensive standard data represent the efficacy of the proposed method. The paper is well-written and easy to follow.
- 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 performance of the whole pipeline depends on the first step of using IntroVAE model. Quantifying the effect of this step separately will further improve the strength of the paper.
- 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
All the necessary checklist for the resproducibilty are provided.
- 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
• Page2, “anomaly detection based methods are unsuitable …” This statement needs to be reformulated. In fact, there have been quite many studies in the field of unsupervised anomaly segmentation of brain tumors. However, most of them relied on single modal MR images such as FLAIR in which at least a part of the tumor appears with hyperintensity patterns. Although the segmentation problem is simplified in such works, some interesting results have been reported. Therefore, this is recommended to reformulate the statement. • Page2, second paragraph, line8: “in this was, tumor regions..” this sentence is a bit confusing and is hard to follow. Please rephrase it. • Page2, second paragraph, line11: “tumor brain images” please reformulate this and the rest in the whole manuscript to “brain tumor images” • Page3, first paragraph, the last three lines: Does this framework function as the same for the testing phase? If so, how the pathological slices will be defined? Otherwise, if in the testing phase, all the slices are analyzed, this should be briefly explained to avoid confusion. • Section 2.1: It was stated that T1 sequence is the most commonly used brain imaging modality, and therefore, the normal appearance model was developed based on T1 images. In addition to the conventionality, it would be interesting to investigate how other sequences perform in capturing the appearance of healthy images. • Another important comment: Since the performance of the whole pipeline depends heavily on the IntroVAE, It will be of great interest to quantify the performance of the IntroVAE iteself. For example, some post processing steps can be conducted on the residual images between the original and the output of the IntroVAE and compare against the segmentation labels. This would be beneficial to quantify the functionality of this piece of the model. • Page 6, line2: Please describe briefly how histogram matching was performed (e.g, one image per modality was randomly chosen and used as reference or ….) . • Regarding the external comparison against nnUNet: please specify if the nnUnet is trained on 2D or 3D. • It will be very good to add another column in table 1 and show the best performance reported over the same BraTS dataset. • Page8, line9: briefly describe how the attention maps were calculated and visualized.
- 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
8
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposed an interesting novel idea to improve the segmentation accuracy. The method sound both theoretically and experimentally. It was well described and the experiments are sufficient to justify the model efficacy. Quantification over extensive experiments are done quite well.
- 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
8
- [Post rebuttal] Please justify your decision
The proposed method sounds technically and the response of the authors to the comments in the context of 4000 characters seems to be sufficient. However, I still encourage the authors to follow the comments in left in the phase1 of the review process: • Another important comment: Since the performance of the whole pipeline depends heavily on the IntroVAE, It will be of great interest to quantify the performance of the IntroVAE iteself. For example, some post processing steps can be conducted on the residual images between the original and the output of the IntroVAE and compare against the segmentation labels. This would be beneficial to quantify the functionality of this piece of the model.
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 paper proposes a brain tumor segmentation network by adopting normal brain images as a reference, which are compared with tumor brain images in the learned feature space. The main contribution is to use normal brains to help the network in contrasting the tumor to the normal region. This idea seems novel and also allows to increase the performance of tumor segmentation. However, the paper should be improved by addressing the following issues:
- The authors claim to use contrastive learning to align the features, but the employed method is SimSiam, which is a non-contrastive learning method. How to explain it?
- Since the performance of the whole framework is highly dependent on the IntroVAE, the authors should give more details on how to train it.
- Capturing the appearance of normal-looking brains does not necessarily require all four MR sequences. It is true, but why is the T1 MRI sequence chosen as the reference, why not Flair?
- In BraTS, a brain tumor can be divided into 3 major regions: enhancing tumor (ET), Tumor Core (TC), and Whole Tumor (WT). Can the proposed method handle multi-class segmentation? If so, why not segment the three tumor regions?
- The authors should explain if the segmentation backbone and Normal Appearance Network were trained from scratch or they are pre-trained models.
- 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).
6
Author Feedback
We thank all reviewers for giving us many valuable comments. We appreciate that they agree on the strengths of our work in terms of novel, interesting, and well-written. Our detailed responses are given below:
1) Why use nnUNet as the only SOTA method, and the improvement of the proposed method is not large? A: We had tested DenseUnet, ResUnet, VNet, attention UNet, and nnUnet, which are usually regarded as SOTA methods and open-source. The resulting average Dice indices of BraTS19 are 0.894, 0.862, 0.875, 0.866 and 0.905, respectively. So we chose nnUnet, which has the highest Dice index of all SOTA methods tested in our experiment. That is also the reason that the improvement of our method seems not large comparing with nnUnet, but with statistical significance (p<0.05 as shown in our paper).
2) The SimSiam is a non-contrastive learning method. A: Though SimSiam does not explicitly use negative samples, it can be thought as BYOL, which is a kind of implicit contrastive learning method, without momentum encoder [C. Zhang, et al., How does SimSiam avoid collapse without negative samples? A unified understanding with self-supervised contrastive learning, ICLR 2022].
3) How was the training and performance of introVAE? A: All normal brain images of IXI dataset (i.e., training dataset of introVAE) went through the same pre-processing as the tumor brain image dataset to reduce possible distribution shift. The encoder and decoder of introVAE were trained iteratively with the learning rates of 1e-4 and 5e-3, respectively. The batch size and max epochs were set to 120 and 200. Other hyper-parameters, like weight terms in the loss function, were the same as the original paper. We think the main concern of reviewers is on the sensitivity to the quality of reconstructed “normal” brain images, and we had conducted an experiment, where normal brain images from healthy people (i.e., less consistent with the tumor brain images comparing with that using introVAE) were adopted instead of the reconstructed images. The average Dice indices are 0.9155 (BraTS19) and 0.8949 (in-house), which are relatively stable.
4) Why segment whole tumor not subregions? A: We used both in-house and public datasets in our experiment. The in-house dataset only has the manually labeled whole tumor as the ground truth. To utilize both datasets, the proposed method focused on whole tumor segmentation. For sub-region segmentation, the CLFC module in the proposed method should be extended by defining feature type for each sub-region, and contrastive learning with negative samples (each sub-region vs. normal region) has to be used instead of SimSiam.
5) Comparing with other contrastive learning methods. A: The objective of our experiment is to prove that using normal brain images as reference can boost the segmentation of tumor brain images. Comparing different contrastive learning methods goes off the objective and can be regarded as an interesting auxiliary experiment.
6) Why Baseline1, which is general Unet, outperforms nnUnet? A: The nnUnet also adopts generic Unet, but it strengthens data pre-processing and training parts, and both parts were also strengthened in the same way in the baseline methods. So similar results can be achieved.
7) Overfitting problem of five-fold patient-wise cross validation. A: Each time, the data was divided into training, validation and testing data. The validation data was used to lock the best model, and the reported results were based on the testing data. So overfitting can be avoided.
8) Why is the T1 MRI sequence chosen as the reference, why not Flair? A: Our study is in the context of clinical routine, where T1 is the most common modality for the normal brain, and multi-modalities are only adopted for the lesion brain.
9) If the segmentation backbone and Normal Appearance Network were trained from scratch or pre-trained? A: they were trained from scratch.
The source code will be released with the paper.
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.
The paper proposes a brain tumor segmentation network by adopting normal brain images as a reference, which are compared with tumor brain images in the learned feature space. The main contribution is to use normal brains to help the network in contrasting the tumor to the normal region. This idea is new and interesting, and also allows to increase the performance of tumor segmentation evaluated on two datasets which include the in-house dataset and BraTS2019 dataset. My proposition is therefore “acceptance”.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- 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).
3
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.
The authors were responsive to some critiques but could have provided a more satisfactory response to other concerns. Particularly, the reasoning for binary segmentation given separate analyses conducted for the two datasets were not convincing, and the analysis could be made more rigorous to further evaluate the usefulness of IntroVAE. However, the authors sufficiently addressed all other concerns including the choice of SimSIAM and providing more details of the methodology. The methodological contributions using a contrastive learning approach to learn from normal brains to segment tumors, together with a reasonably rigorous analysis done using two different datasets, makes for a useful contribution to the MICCAI community.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- 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).
2