List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Mingxuan Lu, Tianyu Wang, Hao Zhu, Mian Li
Abstract
Placenta Accreta Spectrum (PAS) can lead to high risks like excessive blood loss at the delivery. Therefore, prenatal screening with MRI is essential for delivery planning that ensures better clinical outcomes. For computer-aided PAS diagnosis, existing work mostly extracts radiomics features directly from ROI while ignoring the context information, or learns global semantic features with limited awareness of the focal area. Moreover, they usually select single or few MRI slices to represent the whole sequences, which can result in biased decisions. To deal with these issues, a novel end-to-end Hierarchical Attention and Contrastive Learning Network (HACL-Net) is proposed under the formulation of a multi-instance problem. Slice-level attention module is first designed to extract context-aware deep semantic features. These slice-wise features are then aggregated via the patient-level attention module into task-specific patient-wise representation for PAS prediction. A plug-and-play contrastive learning module is introduced to further improve the discriminating power of extracted features. Extensive experiments with ablation studies on a real clinical dataset show that HACL-Net can achieve state-of-the-art prediction performance with the effectiveness of each module.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43990-2_29
SharedIt: https://rdcu.be/dnwLE
Link to the code repository
https://github.com/LouieBHLu/HACL-Net
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a Hierarchical Attention and Contrastive Learning Network (HACL-Net) for MRI-Based Placenta Accreta Spectrum Diagnosis. It mainly has three key points: slice-level attention for context feature learning, patient-level attention for slice feature aggregating, and contrastive loss for discriminative feature learning.
- 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.
- This paper is clear-written, and the experiments are also sufficient.
- considering the context feature and multi-slice problems in MRI-based placenta accreta spectrum diagnosis.
- The patient-level attention-based pooling seems interesting, using trainable parameters to learn the importance of different slices in the MRI images
- 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 whole method seems a little less innovative, the attention methods and the contrastive learning method are all from the known methods.
- Some of the statements or some ideas in the paper are not clear, or not straightforward. For example, why did this paper needs mask branch to detect the ROI region?
- 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
good
- 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
- Even though the experimental results are very impressive, the novelty of this method needs to improve.
- In Trunk branch, the top convolutional layer is for shallow feature extraction. But there is no experiment to show it is effective to use the shallow features for slice-level attention. How about using features in the very beginning or after the ResNet-18?
- How much positive and negative samples in the dataset?
- 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
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The experimental results are pretty good.
- 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
A hierarchical attention mechanism is proposed to solve the prediction problem of placenta accreta spectrum disorder, which achieves the state-of-the-art compared to existing methods and traditional approaches. Although the method is very simple, it has some practical potential.
- 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 structure of the manuscript is well organized.
- 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.
-
In the “Introduction” setction, the authors should use pictures to illustrate the background of the problem, for more clarity.
-
The authors claim that the choice of pre-trained baseline networks has no significant impact, but their results are not the same according to the Table 1. Has the results in Table 1 been statistically tested?
-
Lack of qualitative comparison or presentation of imaging results.
-
- Please rate the clarity and organization of this paper
Satisfactory
- 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 data and code used in this manuscript are not accessible, which may affect the 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 the “Introduction” setction, the authors should use pictures to illustrate the background of the problem, for more clarity.
-
In the “Results” section, the authors should add the final imaging results.
-
- 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 structure of the manuscript is well organized.
- 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
Most of the concerns have been clarified, so my opinion is to accept it but weakly.
Review #4
- Please describe the contribution of the paper
This paper proposed a novel end-to-end HACL-Net for MRI-Based placenta accreta spectrum diagnosis. The proposed network employs slice-level spatial residual attention to extract context-aware deep semantic features and combines them using patient-level attention-based pooling to generate a comprehensive patient representation. Additionally, a contrastive loss is included to enhance the discriminative power of the learned patient-level features.
- 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.
(1) Slice-level attention module is proposed to extract context-aware deep semantic features.
(2) The proposed network designed a patient-level attention module for creating a comprehensive representation of each patient.
(3) To improve the discriminatory power of patient-level features, they also introduce a contrastive loss.
(4) The paper is 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.
(1) For table 1, it seem FPN achieved the best result, why do you still choose U-net in the slice-level attention?
(2) More implementation details should be provided to enable easy reproduction of the work. For instance, the number K_i may vary among different patients and with a batch size of 3, it is unclear how this problem is handled. One possibility is that the output of the slice-level attention is obtained through a for-loop in each batch.
(3) For binary classification, why are the results produced by the softmax activation function and the sigmoid activation function so different?
- 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 data is not publicly available, therefore releasing the code would enable others to reproduce this work.
- 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
Please address the problem mentioned in weaknesses section.
- 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
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Based on main strengths and main weaknesses of the paper.
- 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.
The study focuses on Placenta Accreta spectrum diagnosis on MRI using hierarchical attention and contrastive-learning network.
Strengths of the study include
- novel use of multi-instance learning in the context of placenta accreta spectrum diagnosis.
- nice use of labels to segment the placenta and create a pretrained network for the classification task.
Weaknesses include:
- Its unclear what there the shortcomings of papers mentioned in references 24/26 or 8/6/17, that the present paper is seeking to address. As such also, the rest of the introduction could benefit from rewording to increase clarity.
- Table 2 presents the comparison between existing methods, radiomic methods and the proposed one. The sensitivity of the alternative strategies is really reduced, ranging from 0-at most 35, causing some concerns about the training of such methods.
Please address these comments as well as those of the individual reviewers.
- It remains unclear what is the distribution of Placenta Accreta within the cohort? What is the distribution in the training data? Please also comment regarding the general pregnant population prevalence.
- Please provide what are the novel contributions of this study.
Author Feedback
Thanks for your careful and valuable comments. We will address the questions by category.
1) Dataset details
The cohort consists of 359 subjects selected according to the standards of suspicious PAS (e.g., over three times of pregnancy already; placenta previa) from 2017 to 2022. For the clinical outcomes, 101 subjects show PAS (positive) during the delivery and the remaining 258 do not (negative). The training data consists of 80 positive subjects and 206 negative subjects. The prevalence of PAS in general pregnant population is around 0.2% worldwide and 0.14% in the studied hospital.
2) Novel contributions of paper
Compared to existing literature, the proposed method effectively addresses the limited portion of focal area by leveraging both global context information and ROI-specific information. The network design also utilizes the diversity of multiple MRI slices, aligning with realistic PAS diagnosis by physicians.
Regarding technical details, the pre-trained placenta segmentation network is modified and used as the mask branch of spatial residual attention. Multi-instance learning and contrastive learning are incorporated to learn discriminative characteristics of each patient before classification.
3) Network design
[R1: Q6.2] To justify this choice, comparison results of the three options will be added in the supplementary. Previous experiments show that using the original image as the trunk branch features leads to slice-level attention operating on pixel values instead of structured representations with semantic meanings. Using deep features at the end of ResNet-18 (size of 512) for attention does not effectively enhance ROI details since high-level features lack localized information.
[R2: Q3.2 & R4: Q3.1] The results in Table 1 are averaged over five random train-test procedures. Standard deviations will be added. While there are indeed differences among the compared networks, they are not considered as significant since the segmentation part is not our main contribution. Employing U-Net in the proposed method is based on its fast speed and stable performance, as the common baseline in medical image analysis. The diagnosis performance can potentially be improved with other segmentation networks.
[R4: Q3.3] The sigmoid activation function in the ablation study is to show the general confidence values of diagnosis and to provide insights from the trend with changing thresholds. Results differ from softmax because softmax does not have an explicitly set decision threshold, unlike sigmoid. The predicted scores between positive and negative patients can be close for hard samples, leading to large deviations in sigmoid results.
4) Clarity issues
The ambiguous parts will be clarified in the modified paper, including the use of references in the introduction.
[R1: Q3.2] The mask branch is part of the spatial residual attention, which enhances focal area information and suppresses noise for trunk branch features.
[R2: Q3.1] A figure comparing MRI slices between a pair of patients with and without PAS will be included within the page limit, highlighting the placenta area and abnormal symptoms.
[R2: Q3.3] Due to page limit, visual diagnosis results are not included in the original paper. However, one positive case will be selected to visualize the feature maps extracted by proposed method, other methods and the variants of proposed method in the ablation study. Fig. 4 may be moved to the supplementary.
[R4: Q3.2] Implementation details will be added in the supplementary. The entire code will be released with README. A minibatch of three patients (random, positive and negative) are selected only during training. Slice-level attention is implemented for each MRI slice of a patient through a for-loop. Distinct K_i values are allowed in that K_i number of slice-wise features is aggregated through patient-level attention. For contrastive loss, the aggregated features of patients in a minibatch are compared.
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.
This is a novel approach for Placenta Accreta spectrum diagnosis on MRI using hierarchical attention and contrastive-learning network. This work has great clinical significance. I recommend the acceptance of this study.
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.
Technical novelty is a bit weak but OK. One reviewer increased the score after reading the rebuttal. Overall the paper is at the acceptance level with all the three reviewers gave it “weak accept”.
Meta-review #3
- 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.
Overall, the authors have addressed the concerns from the reviewers, and the experiments are also sufficient to validate the effectiveness of the proposed model. Thus, I recommend accepting this paper.