List of Papers Browse by Subject Areas Author List
Abstract
Current medical image classification efforts mainly aim for higher average performance, often neglecting the balance between different classes. This can lead to significant differences in recognition accuracy between classes and obvious recognition weaknesses. Without the support of massive data, deep learning faces challenges in fine-grained classification of fatty liver. In this paper, we propose an innovative deep learning framework that combines feature decoupling and adaptive adversarial training. Firstly, we employ two iteratively compressed decouplers to supervised decouple common features and specific features related to fatty liver in abdominal ultrasound images. Subsequently, the decoupled features are concatenated with the original image after transforming the color space and are fed into the classifier. During adversarial training, we adaptively adjust the perturbation and balance the adversarial strength by the accuracy of each class. The model will eliminate recognition weaknesses by correctly classifying adversarial samples, thus improving recognition robustness. Finally, the accuracy of our method improved by 4.16%, achieving 82.95%. As demonstrated by extensive experiments, our method is a generalized learning framework that can be directly used to eliminate the recognition weaknesses of any classifier while improving its average performance. Code is available at https://github.com/HP-ML/MICCAI2024.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3325_paper.pdf
SharedIt Link: https://rdcu.be/dVY85
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72378-0_7
Supplementary Material: N/A
Link to the Code Repository
https://github.com/HP-ML/MICCAI2024
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Hua_Robustly_MICCAI2024,
author = { Huang, Peng and Hu, Shu and Peng, Bo and Zhang, Jiashu and Wu, Xi and Wang, Xin},
title = { { Robustly Optimized Deep Feature Decoupling Network for Fatty Liver Diseases Detection } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15001},
month = {October},
page = {68 -- 78}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors design a feature-decoupled network framework for the task of detecting fatty liver in ultrasound images, in which an adversarial-based robust learning method is novelly introduced. Numerous experimental results demonstrate the effectiveness of the authors’ proposed method.
- 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 authors design a practical feature decoupling-based network framework for the task of fatty liver classification.
- the novel introduction of adversarial-based robust learning method is applicable to the classification task.
- the proposed method in the article has some generalization for medical image classification tasks.
- 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 authors design a practical feature decoupling-based network framework for the task of fatty liver classification.
- the novel introduction of adversarial-based robust learning method is applicable to the classification task.
- the proposed method in the article has some generalization for medical image classification tasks. ● Weaknesses.
- Lack of clarity: the article does not clearly state the mechanism of action of the adversarial-based training phase. A large number of variables are introduced in the method introduction section and the block diagram of the algorithm but lack of specific explanations, e.g., what Lc, Ls, and Lat stand for and what they effect respectively.
- The effectiveness of contrastive learning is questionable. The methods section is not clear about contrastive learning. In addition, in the ablation experiments, the addition of contrastive learning only resulted in a very weak performance improvement.
- how more robust decision bounds can be achieved by adding perturbations and classifying them in contrast learning. As I understand it, the classifiers in this section are not shared with the classifiers for fatty liver classification. The gradient of the classification loss Lat simply conducts to the decoupled to features. Further explanation is needed.
- The training process of the model is complex and involves feature decoupling and adversarial learning. The authors should clearly articulate the process of the inference stage to avoid readers’ concerns about the efficiency of the model inference.
- In order to facilitate readers’ understanding of the importance of feature decoupling in medical ultrasound image analysis, the authors should cite more related studies as the background of the method.
- In order to demonstrate the effectiveness of feature decoupling, proper visualization of feature maps is necessary, if the authors can provide it.
- 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 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
as shown in the weaknesses
- 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?
The study has some novelty and can bring some inspiration to the research community however some questions remain to be further explained.
- 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
Weak Reject — could be rejected, dependent on rebuttal (3)
- [Post rebuttal] Please justify your decision
The methodology is certainly innovative. But the authors did not do a good job of explaining why it works.
Review #2
- Please describe the contribution of the paper
A deep learning framework that combines feature decoupling and adaptive adversarial training has been proposed for improving the robustness of fatty liver recognition.
- 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.
- Well-motivated introduction section
- An ablation study has been conducted
- Decent performance improvement
- 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.
- Certain statements in the abstract lack clarity
- Shallow explanation of data acquisition
- Parameter optimization procedure has not been discussed.
- 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 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
- Abstract: “deep learning is difficult to classify…”- does the sentence correctly deliver the intended idea? Do you mean “deep learning undergoes difficulty…”?
- Abstract: “Finally, the average performance…”- please mention the specific metric for this quantitative result.
- Algorithm 1: Please make the caption more informative.
- Equations 2-4: How do you tune/optimize the hyperparameters?
- Dataset: Please extend the data collection section to include patient recruitment procedure/criteria, ultrasound machine and imaging parameters, etc.
- Training: Please include the training process and the associated parameters.
- Table 1: Please use different markings for the best and worst results.
- Table 2: Why does SENet + ICFDNet reduce the accuracy?
- 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?
Factors:
- Not confident about reproducibility
- Decent performance improvement
- Well-conducted ablation study
- 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 authors have addressed my reproducibility-related concern and earned my vote for the acceptance of this paper.
Review #3
- Please describe the contribution of the paper
1)A new deep learning model, Iterative Compression Feature Decoupling (ICFDNet), is proposed for feature decoupling.ICFDNet improves the sample utilisation rate and ensures that the information is fully extracted and used. 2)Colour space transformation and selection strategies are used to enrich the dataset and minimise noise interference in model optimisation. 3)Adversarial training is used to find the weakness of the model and use it to generate more robust decision boundaries. In addition, the perturbation and adversarial strengths are adaptively adjusted according to the accuracy of each category, achieving an overall improvement in model capability.
- 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 authors provide thorough and detailed explanations of the proposed methodology, ensuring that readers can fully understand the underlying concepts and implementations. This clarity of presentation demonstrates the authors’ deep understanding of the subject matter and enhances comprehensibility for the target audience. A key strength of this paper is its comparison with previous approaches in the field, effectively demonstrating the improvements and advantages offered by their work, thus highlighting its potential impact on the field. The paper concludes with a well-organised and succinct summary of the key findings, contributions and implications of the research. This section effectively synthesises the key points of the paper, enabling the reader to understand the significance of this work and its potential to advance the current state of knowledge in the field.
- 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) ICFDNet: figure 3 in 2.1 is unclear in the description about the legend and lacks a detailed description of the network decoupling process. (2) Inadequate description of previous approaches: although the paper compares the proposed approach with earlier approaches in the field, the authors do not provide enough information about these previous approaches. (3) In the experimental section of Chapter 3, most of the algorithms compared in Table 1 are relatively old. (4) The comparison of ablation experiments for ViT is missing in Table 2.
- 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 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
- Add a detailed description of the network decoupling process in 2.1 and improve the legend description in Figure 3.
- A short but informative overview of earlier approaches would be valuable in helping the reader understand the limitations of existing approaches and the reasons behind the development of the proposed method. This additional context would further emphasise the novelty and potential benefits of the authors’ work.
- A comparison of MICAAI’s algorithms from the last two years should be added.
- Insufficient description of the dataset: the authors do not provide an adequate description of the dataset used in the study, which may hinder the reader’s understanding of the context and applicability of the proposed methodology. Lack of description of dataset partitioning
- 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 paper provides a thorough and detailed explanation of the proposed methodology, ensuring that the reader fully understands the underlying concepts and implementation. However, the ablation experiments are incomplete due to the lack of detailed descriptions of some of the components, as well as the lack of comparisons with the state-of-the-art algorithms in the comparison experiments.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Author Feedback
We thank all reviewers for their thoughtful review and constructive comments. The following are our responses to the major critiques. 1) Expression and marking(R3Q1,2,3&7): We will revise Abstract to eliminate ambiguous expressions. The caption of Alg. 1 will be updated to “Feature Decoupling and Adaptive Adversarial Training Optimization Process.” The best and worst results in Table 1 will be marked differently. 2) Hyperparameters(R3Q5&6): The hyperparameter values in Eq. 1-4 are: λ1=0.2, λ2=0.4, λ3=1, ε=2/255, β=1. Additional hyperparameters will be detailed in the experimental section. 3) SENet+ICFDNet accuracy(R3Q8): A key challenge in fatty liver classification is the varying difficulty of recognizing different classes. This can make the normal training process more likely to optimize towards simpler classes, resulting in unbalanced performance. SENet shows this issue, while ICFDNet mitigates it, ensuring balanced performance across classes. The final result improves both balance and average accuracy. 4) Data collection and partitioning(R3Q5, R4Q4): Data were collected from patients aged 65+ diagnosed with fatty liver. Due to the extended collection period, ultrasound machine parameters varied based on the patient. While we cannot specify the exact settings, we ensured that two physicians confirmed the diagnostic results for each sample. For the experiment, the dataset was split into training, validation, and testing sets in a 7:1:2 ratio. 5) Table 2 missing ViT(R4): Due to memory requirements, our device couldn’t support ViT integration into our method. When testing ViT, we split the decoupling and the adversarial training(AT) into two sequential processes. ViT achieved 69.99% accuracy. Although it has a significant improvement, it was not tested under the same conditions. So, it is not listed in Table 2. 6) Decoupling process(R4Q1): The decoupling process involves two parallel branches: one extracting specific fatty liver features and the other extracting common ultrasound features using ICFDNet with unshared parameters. Special features are supervised with BCELoss after MLP, while common features are supervised using image reconstruction loss after a convolutional layer. 7) Algorithm selection in Table 1(R4Q3): Our requirement for comparative methods is that they are highly generalizable. Most existing works use transformers and CNNs as their main structures, then integrate residual structures and attention mechanisms for specific optimizations. Following your comment, we tested the FairAdaBN(MICCAI 2023), which achieved 78.41% accuracy, which is lower than the best result of 82.95%. 8) Related work(R4Q2, R5Q5): We will add more previous related work in Introduction. 9) Figures and visualizations(R4Q1, R5Q6): At this stage, we can only provide text and can’t offer visual results. If space permits, we will include visualizations in the manuscript or release them as open-source material. We will add more detailed legends to the figures. 10) Further explanation for AT and its effectiveness(R1Q6, R5Q2&3): Images are processed by ICFDNet to get two decoupled features. These features are concatenated with the Y-channel and sent to the classifier for further training. Adversarial loss affects the entire training process by maximizing the KL divergence between adversarial and clean samples. AT identifies and optimizes classifier weaknesses by causing errors and adjusting its adversarial strength based on the accuracy of each class. Experimental results show all classifiers improved, with the best achieving a 3.4% increase, demonstrating our method’s generalizability. Additionally, our method only increases training consumption without adding inference complexity. 11) Inference process(R5Q4): Images are fed to two ICFDNet models in parallel. The decoupled features are concatenated with the Y-channel and fed into the classifier for inference results. 12) Reproducibility(R3, R4, R5): The code and data will be shared via Github.
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’
After the rebuttal, this paper received an “Accept,” “Weak Accept,” and “Weak Reject.” The authors addressed most concerns raised by the reviewers. R5 stressed the clarity of the method, but based on my reading of the authors’ rebuttal, the answers are clear, and the authors committed to publishing the code upon acceptance. Therefore, I suggest an “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).
After the rebuttal, this paper received an “Accept,” “Weak Accept,” and “Weak Reject.” The authors addressed most concerns raised by the reviewers. R5 stressed the clarity of the method, but based on my reading of the authors’ rebuttal, the answers are clear, and the authors committed to publishing the code upon acceptance. Therefore, I suggest an “Accept.”
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