List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Jiacheng Ruan, Mingye Xie, Jingsheng Gao, Ting Liu, Yuzhuo Fu
Abstract
Transformer and its variants have been widely used for medical image segmentation. However, the large number of parameter and computational load of these models make them unsuitable for mobile health applications. To address this issue, we propose a more efficient approach, the Efficient Group Enhanced UNet (EGE-UNet). We incorporate a Group multi-axis Hadamard Product Attention module (GHPA) and a Group Aggregation Bridge module (GAB) in a lightweight manner. The GHPA groups input features and performs Hadamard Product Attention mechanism (HPA) on different axes to extract pathological information from diverse perspectives. The GAB effectively fuses multi-scale information by grouping low-level features, high-level features, and a mask generated by the decoder at each stage. Comprehensive experiments on the ISIC2017 and ISIC2018 datasets demonstrate that EGE-UNet outperforms existing state-of-the-art methods. In short, compared to the TransFuse, our model achieves superior segmentation performance while reducing parameter and computation costs by 494x and 160x, respectively. Moreover, to our best knowledge, this is the first model with a parameter count limited to just 50KB. Our code is available at https://github.com/JCruan519/EGE-UNet.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43901-8_46
SharedIt: https://rdcu.be/dnwDU
Link to the code repository
https://github.com/JCruan519/EGE-UNet
Link to the dataset(s)
ISIC2017: https://challenge.isic-archive.com/data/#2017
ISIC2018: https://challenge.isic-archive.com/data/#2018
Reviews
Review #3
- Please describe the contribution of the paper
EGE-U Net, for Efficient group Enhanced UNet is proposed for segmenting with transformers, biomedical images in mobile applications. They use a Group multi-axis Hadamard product attention mechanism (GHPA) to extract information with a linear complexity and a Group Aggregation Bridge module (GAB) to fuse information at each decoder stage. They compare the SATO (TransFuse which uses a CNN and Vision Transformer dual path) and their method on two datasets outperforming it in performances by lightning parameters and costs. It may be the first architecture of segmentation with 50kb parameters.
- 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.
Light. The architecture proposed is lightened and represents only a few costs. Validation. The validation has been made on two datasets with a literature comparison and various metrics. Extendability. This architecture could be used on other biomedical datasets.
- 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.
Limitation in application use. In the summary the mobile health application is mentioned and is not applied inside this article. It may be used in a perspective sentence? And the justification may be more general in the abstract?
Limitation in justification of the Lambda used for the loss. How has the loss been chosen?
- 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
Reproducible.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
In Figure 1, the GFLOPs are not in the same gradation colors, why? Is there another difference than the two datasets ? For homogeneisation the authors should keep the same color gradation for the same metrics.
In Figure 3, the black arrow should be in colors such as the others for figure homogenisation.
Why do not use a λi between 1 and 0.5 in other steps? To justify in the text.
Could we have the original size in the description of the dataset? Why do not partitioning the dataset in test, train and validation? To justify why only test and train.
In future comparisons, could we have also the accuracy and F1-score (specificity, sensibility etc.), two metrics are minimum.
The best results in the ablation Table should be highlighted in bold (a)Baseline + GHPA, b)first column DW, second column multi-axis c) depends… and the text must be linked to this part for the c).
In Figure4 : it lacks the false positive and false negative zones in an other color or an arrow which could help to see the differences between the ground truth and the various architecture segmentation results.
It may be interesting to know how many epochs have been used to compute the other architectures or the computation time even if the GFLOP are given.
- 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?
cf the minor weaknesses found and the main contribution.
- 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 #2
- Please describe the contribution of the paper
An novel EGE-Unet model was proposed for image segmentation. The model has the smallest set of parameters and achieved the best segmentation results compared with existing mdoels on 2 public 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.
An novel EGE-Unet model was proposed for image segmentation. The model has the smallest set, runs fatest, of parameters and achieved the best segmentation results compared with existing mdoels on 2 public datasets.
- 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.
No major weakness
- 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 code is avaiable (along with the submission).
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
An novel EGE-Unet model was proposed for image segmentation. The model has the smallest set of parameters and achieved the best segmentation results compared with existing mdoels on 2 public datasets. Overall, it is a well written paper, and is acceptable.
- 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?
Solid evaluation results indicated the significant improvement in this field.
- 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 #1
- Please describe the contribution of the paper
The EGE-UNet is an efficient network which used for skin lesion segmentation. It has two new module called GHPA and GAB with the former efficiently acquiring and integrating multi-perspective information and the latter accepting features at different scales, along with an auxiliary mask for efficient multi-scale feature fusion.
- 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.
- Introduction of the paper is well written and provides a good literature survey of the field.
- The contributions are stated clearly.
- The proposed network is very useful for the mobile health applications.
- 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.
More experiments should be supplement in section 3.
- 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 authors provided the codes and can be reproducibility.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- In equation 1, what does the mean of y and y^? This loss function should consider the different i in the right of equals sig.
- Please supplement the ACC, SPE, SEN in the experiments.
- 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 EGE-UNet can reduce parameter to approximately 50KB, while keeps a good mIoU and DSE.
- 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
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.
A novel approach is proposed and the paper is well-written with clearly-stated contributions. The approach also achieves improved results over competing models. The wider applications of this work should be discussed in some more detail and some more justification of the notation used in the methodology would be beneficial. It would also be useful to expand the evaluation to include further relevant metrics that are used in the literature.
Author Feedback
Respond to Reviewer #1 Thank you for your recognition and suggestions on this work! Q1: The mean of equation 1. A: In equation 1, y represents the ground truth, y^ represents the model’s prediction, and i=0,1,2,3,4,5.
Respond to Reviewer #2 Thank you for your recognition of this work!
Respond to Reviewer #3 Thank you for your recognition and suggestions on this work. Your suggestions are very insightful! Q1: Why do you use different colors in Figure 1? A: Different colors are only used to represent two different datasets.
Q2: Why do not use a \lambda_i between 1 and 0.5 in other steps? A: As mentioned in the paper, we used the technique of deep supervision, which is a common technique in medical image segmentation. Based on previous experience, we set \lambda_i to 1, 0.5, 0.4, 0.3, 0.2, 0.1, and the deep supervision technique is only used during training. If \lambda_i in other steps is set to be too large, it may cause the model to consider the intermediate output masks to be equally important as the final output mask, which may affect the final segmentation performance. Further ablation experiments are needed for specific verification, which is not the focus of this work.
Q3: Regarding the dataset. A: For the ISIC17 and ISIC18 datasets, the original size of the images is not uniform. For example, some images have a size of 2848×4288, while others have a size of 768×1024, and there are still other images with different sizes. Additionally, it is worth noting that, in order to make a fair comparison, our work followed the experimental settings of MALUNet [Ruan J, Xiang S, Xie M, et al. MALUNet: A Multi-Attention and Light-weight UNet for Skin Lesion Segmentation[C]//2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2022: 1150-1156.].
Respond to META-REVIEWER #2 Thank you for your recognition and suggestions on this work, which are very insightful!