List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Mehmet Saygın Seyfioğlu, Zixuan Liu, Pranav Kamath, Sadjyot Gangolli, Sheng Wang, Thomas Grabowski, Linda Shapiro
Abstract
We propose a novel framework for Alzheimer’s disease (AD) detection using brain MRIs. The framework starts with a data augmentation method called Brain-Aware Replacements (BAR), which leverages a standard brain parcellation to replace medically-relevant 3D brain regions in an anchor MRI from a randomly picked MRI to create synthetic samples. Ground truth hard" labels are also linearly mixed depending on the replacement ratio in order to createsoft” labels. BAR produces a great variety of realistic-looking synthetic MRIs with higher local variability compared to other mix-based methods, such as CutMix. On top of BAR, we propose using a soft-label-capable supervised contrastive loss, aiming to learn the relative similarity of representations that reflect how mixed are the synthetic MRIs using our soft labels. This way, we do not fully exhaust the entropic capacity of our hard labels, since we only use them to create soft labels and synthetic MRIs through BAR. We show that a model pre-trained using our framework can be further fine-tuned with a cross-entropy loss using the hard labels that were used to create the synthetic samples. We validated the performance of our supervised pre-training with synthetic MRIs plus a fine-tuning framework in a binary AD detection task against both from-scratch supervised training and state-of-the-art self-supervised training plus fine-tuning approaches. Then we evaluated BAR’s individual performance compared to another mix-based method CutMix by integrating it within our framework. We show that our framework yields superior results in both precision and recall for the AD detection task.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_44
SharedIt: https://rdcu.be/cVD60
Link to the code repository
https://github.com/aldraus/BrainAwareReplacementsForAD
Link to the dataset(s)
https://ida.loni.usc.edu/login.jsp?project=ADNI&page=HOME
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces an augmentation technique along with a contrastive learning method for pretraining classifiers for Alzheimer’s disease. The study uses augmented labels as “soft labels” for supervised contrastive objective.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper is well organized and it presents the methodology very clearly. It is addressing a significant problem without any unrealistic claims. The authors have formulated the problem and the study’s hypothesis is clear. The proposed method is an incremental improvement on the literature, which is sufficient for publication. The experimental setup is designed to validate the hypothesis of the paper, with sufficient ablation study.
- 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.
I did not see any major weaknesses in the paper. Only a couple of minor points:
- The paper claims: “Also, BAR implicitly forces the model to pay attention to the relationship between medically relevant brain regions, thus making it more clinically relevant.” This statement implies causality. Although the authors have presented some attention maps in the supplementary materials, causal feature learning requires deeper analysis.
- Although the authors have targeted a very specific problem to solve i.e., AD detection in MR, but the general idea behind the method is not specific to AD. It would’ve been better if the authors included other clinical problems of significance.
- 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 dataset is publicly available and the authors will publish their code. So the results should be reliably 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/2022/en/REVIEWER-GUIDELINES.html
I very much enjoyed reading the paper. I suggest the authors expand their scope of their method to other problems as well.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well organized and the objectives clear. It has sufficient technical novelty for MICCAI and the presented method is thoroughly evaluated.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
The authors combine the advantages of relaxed contrastive learning and use case specific data augmentation operations (BAR and BAM) to solve the Alzheimer disease detection. The performance of the presented method is reported on the Alzheimer’s Disease Neuroimaging Initiative 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 paper is well written.
- The BAR and BAM augmentations are novel: BAR is a new version of CutMix which takes benefit from the anatomic properties of the brain while BAM is a new version of Cutout.
- 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 applicability of the BAR augmentation is very limited and can not be generalized to other use cases since it sets hard constraints on the data: alignment and pixel precise information.
- The evaluation is very limited since it does not compare to any of the state of the art methods. Since the method builds on top of ViT, I am missing the comparison with methods using convolutional networks.
- 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 paper reports architectural and experimental setups thoroughly. Thus the results should be with little effort 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/2022/en/REVIEWER-GUIDELINES.html
Since the introduced method is architecture agnostic and convolutional networks are very successful in solving object detection, including results for convolutional networks should make the contribution even stronger. (ViT are successful on image classification but way less efficient on other vision tasks). Also, reporting the performance of previous state of the art methods should help evaluating the importance of the introduced augmentations. An even better scenario would be to take the old state of the art method and include the augmentations in an appropriate way and report the performance.
- 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 limited applicability of the introduced augmentations and the weaknesses in the evaluation push toward not accepting the paper. Minor improvements might lead to upvoting the paper.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
3
- 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 #3
- Please describe the contribution of the paper
The authors propose a new data augmentation strategy for the contrastive learning framework to train a better pre-trained model. Their method produces a great variety of realistic-looking synthetic MRIs with higher local variability compared to other mix-based methods, such as CutMix.
- 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 an interesting idea to produce a great variety of realistic-looking synthetic but not simply mix up some image patches in “CutMix”. “CutMix” has been proved to be effective in natural images. But, in the medical image field, we pay more attention to Interpretability. It is important to input the cases matching the anatomy structure so “CutMix” strategy may not be suitable for the medical images. Meanwhile, they train a supervised contrastive loss with the soft labels and synthetic images, leading to very powerful representation 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.
- They can compare the attention maps (grad-cam) from the different models for some more detailed analysis but not just some metric values. Since they focus on the AD classification, they can compare the performance of replacing brain regions related to with AD (e.g., hippocampus) or other unrelated regions.
- They should compare with more SOTA pretrain methods. In this paper, they just compare with some simple baselines, such as naive contrastive learning, cutmix.
- 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 author will release their codes in github.
- 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
- They can generate the grad-cam maps from the different models for some more detailed analysis but not just some metric values.
- They can compare the performance of replacing brain regions related to AD (e.g., hippocampus) or other unrelated regions.
- They should compare with more SOTA methods of model pretraining, e.g., moco [1].
[1] He, Kaiming, et al. “Momentum contrast for unsupervised visual representation learning.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.
- 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I appreciate that they focus on the problems of “Cutmix” in the medical image field. Their method produces a great variety of realistic-looking synthetic MRIs with higher local variability.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- 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
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.
In general the reviews found the ideas presented in this paper to be novel which I agree with. The concept of brain-aware cut-mix and cut-out augmentations which generate realist augmentations is interesting. While one reviewer saw the narrow focus of the methods to be a weakness, my opinion is that this not a major weakness for the MICCAI conference. Other strengths of the paper were noted to be: good organization and well formulated problem.
The most important weakness of the paper was noted by two of the reviewers to be a lack of comparison to state-of-the-art baselines. While the ablation studies were appreciated, the baselines were thought to be rather simple. Another weakness in the experiments section was the lack of deeper analysis to support statements such as the proposed method paying attention to relationships between medically relevant brain regions.
Overall, I think the paper will be of interest to the MICCAI audience and the strengths of the paper outweigh its weakness.
- 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).
3
Author Feedback
We are very grateful for the reviewers’ insightful feedback and comments. We would like to answer one comment. (R3: They can compare the attention maps (grad-cam) from the different models for some more detailed analysis but not just some metric values.) Due to space limitations, we could not add any other figures to the paper. However, we added the attention maps for both CutMix and BAR to the Supplementary Material.