Paper Info Reviews Meta-Review Author Feedback Post-rebuttal Meta-Reviews

Authors

Hanbo Chen, Kang Wang, Yuyao Zhu, Jiangpeng Yan, Yan Ji, Jingjing Li, Dong Xie, Junzhou Huang, Shuqun Cheng, Jianhua Yao

Abstract

Microvascular invasion (MVI) is a histological feature of hepatocellular carcinoma (HCC). It is the strongest feature related to HCC recurrence and survival after liver resection. However, its diagnosis is time consuming which requires pathologist to examine high resolution histopathological images. A computer aided MVI detection system that improves the diagnosis efficiency and consistency is in demand. There are two challenges in MVI detection (1) MVI is formed by the same type of tumor cells as common cancer tissue (CCT) and (2) both MVI and CCT’s size varies significantly – from a few cells to thousands of cells. Inspired by pathologists’ routine reading procedure, we propose a 3-stage solution composed by cascaded networks to tackle this problem. In this framework, images are first analyzed by pixel-level cancer tissue segmentation, followed by region-level instance feature extraction, and then by slide-level comparison to detect MVI. To reduce inter-stage error accumulation, the system is designed in the way that later stage can learn and correct errors in the previous stages. To effectively conduct slide-level analysis, a novel convolutional graph neural network with short cut (sc-GCN) is proposed to solve the over-smoothing problem in classic GCN methods. Testing results on 90 WSI samples show that the proposed system achieves state-of-the-art performance in MVI detection.

Link to paper

DOI: https://doi.org/10.1007/978-3-030-87237-3_19

SharedIt: https://rdcu.be/cyl9Y

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #1

Please describe the contribution of the paper

This paper proposed a 3-stage cascaded networks to detect MVI regions in WSIs. Patches in WSI are first analyzed by pixel-level cancer tissue segmentation, followed by region-level instance feature extraction, and then by slide-level comparison to detect MVI.
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.

A novel convolutional graph neural network with short cut (sc-GCN) is proposed to solve the over-smoothing problem in classic GCN methods.
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. The dataset in this paper is collected by author not public, so how to achieve state-of-the-art performance on MVI detection?
2. From Table1 we can find the size of MVI and CCT has vary big difference. So if the size of region is the main effective feature to classify MVI should be evaluated in experiments.
3. “To force clear separation be-tween adjacent tissues, we follow previous work [6, 7, 10] and add a third class of 5 μm-thick tissue border into our segmentation target.” The references of 6,7 and 10 are not relative information about this technical.
4. In sc-GCN part, how to get the short-cut method is not mentioned in the paper.
5. In Table2, how to define the area weighted is not described in the paper.
Please rate the clarity and organization of this paper

Poor
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

There are some difficulties to reproduce the paper wcith many key information missed.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html

According to part 4.
Please state your overall opinion of the paper

probably reject (4)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

Many detailed information of this paper is missed.
What is the ranking of this paper in your review stack?

3
Number of papers in your stack

2
Reviewer confidence

Confident but not absolutely certain

Review #2

Please describe the contribution of the paper

The authors present a three-stage framework for microvascular invasion detection in hepatocellular carcinoma. The proposed framework utilizes different kinds of features in each stage and has the ability to correct the mistake from the previous stage. An sc-GCN is proposed to effectively conduct the slide-level analysis.
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.
- A novel cascaded framework for WSI analysis. After conducting a preliminary cancer segmentation in the first stage, a classification network is further utilized both masks and cropped images as input for cancer category classification, which could consider as a refinement of the first stage result with the capability of correct the wrong prediction.
- A novel GCN method to incorporate deep feature and handcrafted feature for slide-level analysis. The authors design a shortcut way in GCN to resolve the over-smooth problem. An ablation study on this component demonstrates the effectiveness of the proposed module.
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.
- Lack of comparison with other WSI analysis methods especially those used GCNs. Actually, the author mentioned in the paper that methods[1-3] also use GNNs to model the neighbouring instance information. However, it lacks the detailed comparison with these related methods, either in description or experiments.
[1] Raju, Ashwin, et al. “Graph Attention Multi-instance Learning for Accurate Colorectal Cancer Staging.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2020. [2] Li, Ruoyu, et al. “Graph CNN for survival analysis on whole slide pathological images.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018. [3] Zhu, Xinliang, et al. “Wsisa: Making survival prediction from whole slide histopathological images.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017.
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 authors provide the training strategy and the detailed structure of sc-GCN. Therefore, it is possible for researchers to reproduce the results.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html
The authors present a three-stage framework for microvascular invasion detection in hepatocellular carcinoma.

In sum, the designing of the proposed method is novel and reasonable, which has considered the characteristics of MVI and the general pathologists’ routine. It starts from a course and easy task and further refines it with different additional information. Experiment results on the HCC dataset is good, with additional ablation studies for the design of each module. The proposed method has the potentially applicable to the diagnosis of other cancers. Yet it still could be improved in different aspects. The main weakness of this paper is the lack of comparison with other GNN based WSI analysis methods [1-3]. Meanwhile, there exists many well-designed GCN method in the research area of graph neural network, e.g., GINConv [4], GraphSage [5], DropEdge[6], especially JK-Net[7] which also adds shortcut in multiple stage. It will be interested to see the comparisons between them.

Minors:
- The resolution of figures is not high enough.
- An ambiguous arrow exists in the figure from supplementary material.
[1] Raju, Ashwin, et al. “Graph Attention Multi-instance Learning for Accurate Colorectal Cancer Staging.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2020. [2] Li, Ruoyu, et al. “Graph CNN for survival analysis on whole slide pathological images.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018. [3] Zhu, Xinliang, et al. “Wsisa: Making survival prediction from whole slide histopathological images.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017. [4] Xu, Keyulu, et al. “How powerful are graph neural networks?.” ICLR 2019. [5] Hamilton, William L., Rex Ying, and Jure Leskovec. “Inductive representation learning on large graphs.” NeurIPS 2017. [6] Rong, Yu, et al. “Dropedge: Towards deep graph convolutional networks on node classification.” ICLR 2020. [7] Xu, Keyulu, et al. “Representation learning on graphs with jumping knowledge networks.” International Conference on Machine Learning. PMLR, 2018.
Please state your overall opinion of the paper

borderline accept (6)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The designing of the proposed method is reasonable, which has considered the characteristics of MVI and the general pathologists’ routine. Yet more experiments need to be done to prove its effectiveness.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

6
Reviewer confidence

Confident but not absolutely certain

Review #3

Please describe the contribution of the paper

A three-stage solution for Microvascular invasion (MVI) detection for hepatocellular carcinoma detection has been proposed. First, they have used U-net with Se-ResNet backbone to classify all candidate (CCT), then Resnet prediction and handcrafted features are concatenated and fed to an sc-GCN network. sc stands for a shortcut that authors added to Graph Conv Nets.
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 problem itself is interesting because there is not that much literature in digital pathology in this regard.
- Blending networks and solving the segmentation in multi-stage is not new, but the first time for MVI detection.
- Idea of Adding one short-cut in GCN has improved the result’s accuracy.
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.
- Sharing the dataset would increase the value of the work much further
- The explanation of three different networks was limited. I could not find the second step details (dual input se-ResNet) even in the supporting materials.
- No mention of cases, just WSIs are mentioned. Two WSIs belonged to the same patient could falsify the results
- Please use the full name of Medical Image Analysis in the references, not Med Img Anal.
- Graphics are inserted in low resolution with kind hard to follow arrows
- Using just one deep network, despite the Intuitive is not enough and testing more backbones like finetuned networks for pathology images would increase the value of the paper
Please rate the clarity and organization of this paper

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

Most part of parameters well addressed in the appendix. However, it needs to add more details for each step. Still hard to implement. For example in the feature extraction phase author states “extract high-order tissue shape features and global texture information” however, it been added later “The predicted probability will be concatenated with hand crafted” Not sure they get one number or feature vector from the deep network.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html
- abrivation for common cancer tissue (CCV) has mentioned in abstract no need to repeat the full name and fabrication in the paper.
- Figure 1 is not easy to follow. The image quality is low, arrow color and text are hard to read.
- Test and train cases are different or no? You need leave one patient (two or more WSIs may belongs to one patient). Having WSIs from one patient will falsify the results.
- I would say a bar chart could represent table one much better.
- Figure 4 is so crowded and I had a hard time to understand it. Bight be better to show in two separate images.
- Explanation about table 2 and 3 difference is not enough. Still not sure why F1 hand features MLP in table 2 (50%) is different from MLP in table 4 (85%). while you stated To isolate the impact of other networks, this analysis is based on the manual annotation and hand-crafted features only
- Please add size of deep feature in the middle stage (if this is just probability then feature is not a sutible term form it)
- in the first place please add full name of se abrivation when you mention to Se-ResNet
Please state your overall opinion of the paper

accept (8)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

Expect 2-3 weakness that are not that much major, the paper was written nicely. I would rather go with accept.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

5
Reviewer confidence

Confident but not absolutely certain

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.

This paper presents a three-stage framework for MVI detection in hepatocellular carcinoma. The application seems new and could be interesting to the community. The main strength lies in the cascaded framework for WSI analysis and incorporating GCN with shortcut to resolve over-smooth problem. In the meanwhile, the reviewers have raised concerns about the difficulty on reproduction, illustration on sc-GCN, lack of direct comparison with other WSI analysis methods using GCN, further explanation on network details, etc. The authors are invited to submit a rebuttal to address these issues.
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).

7

Author Feedback

We like to thank reviewers for their constructive comments and acknowledging the novelty of this work. Detecting MVI on pathology image is an “interesting” (R#4) and “new” (meta-R) problem. The proposed short-cut GCN is “novel” (R#1, R#2) and the 3-stage approach is “novel” (R#2) and “first time for MVI detection” (R#4). Based on reviewers’ comments, we will improve the clarity with details in paper revision.

Below are our responses to major concerns.

Lack of comparison (meta-R, R#1, R#2) In our preliminary experiments, we have compared multiple CNN-based and GNN-based approaches. Our proposed method generates the best final results. Due to paper length limitation, we only keep the comparison that are most relevant to our proposed method, including CNN-based classifier (Table 2, row 2), hand-crafted features (Table 2, row 3), and classic GCN (Table 3). R#1 suggested using size as prediction metric, which is already included in hand crafted features (Table 2, row 3). We tried using size only, but the result is worse since there is overlap between MVI and HCC size (Fig. 1). Some of the WSI method suggested by R#2 are related but cannot be directly used - detecting MVI is an object detection task while method like WSISA is for image classification. We will clarify and add more discussion on these related methods. We value the suggestion of comparing with more GNN structures (R#2) and more CNN backbones (R#4). We will try to include some in the final revision. But it is impossible to do a comprehensive study in each stage with space limit, we have to leave some of them for the future extension work.

Data availability (R#1, R#4) We did not find any public dataset that is suitable for this task. Experimental data is from our collaborator and we are not authorized to make it public in the current stage.

Explain network structure of sc-GCN (R#1, meta-R) and dual input se-ResNet (R#4) Sc-GCN has 2 updates based on GCN network: 1) a short-cut is introduced by concatenating raw node features with hidden variables after convolution; 2) an additional MLP fuses the concatenated features to classify each vertex. We will highlight these differences in Fig. 3 and its legend. “se” in se-ResNet is short for squeeze-and-excitation, which is an operation added after convolution in ResNet. We concatenate segmentation prediction with raw image in the color channel dimension and take it as the input of se-ResNet. Thus, it has dual (two) inputs and is named dual input se-ResNet.

Lack of details to reproduce work (R#1, R#4, meta-R) Some details are missing due to page limitation. Because every stage of the system is optimized for MVI detection and requires detail, it is challenging to explain everything with limited paper length. We are relieved to hear from reviewers that “it is possible for researchers to reproduce the results” (R#2) and “most part of parameters well addressed in the appendix” (R#4). We will revise paper accordingly to include more details and improve clarity. Below are our responses to the remaining unclear part: a) R#1: explain edge segmentation: Instead of binary segmentation, we conduct 3-class segmentation: cancer, cancer edge, others. b) R#1: explain area weight in Table 2: It is defined as the area of the annotation instance. c) R#4: patient separation between train and test set: All WSI images are from separated patients. There is no data leakage in patient level. d) R#4: low image resolution: Resolution is compressed when generating pdf. We will correct it. e) R#4: dimension of stage 2 feature: In stage 2, “high-order tissue shape features and global texture information” refers to the prediction result of image classifier. Size of deep feature is 7 (6 hand crafted features + 1 classifier probability). f) R#4: difference between table 2 and 3: Table 2 predicts MVI based on the predicted segmentation. Table 3 is based on manual annotations. Results in table 3 is better as it eliminates segment error.

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.

Although some details about methodology and experiments are not provided, given the new application to MICCAI research community, I would likely recommend accept. The authors should provide more details regarding the methodology as well as experiments and significantly improve the quality of paper (see comments in R1) in the final version.
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 (best paper in your stack) and n (worst paper in your stack of n papers).

8

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.

A three-stage solution for Microvascular invasion (MVI) detection for hepatocellular carcinoma detection has been proposed. It is a novel and interesting application. And the sc-GCN method is proposed to solve the over-smoothing problem in classic GCN methods. The authors’ rebuttal solves most of the questions of reviewers.
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 (best paper in your stack) and n (worst paper in your stack of n papers).

6

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.

This paper proposed a 3-stage cascaded networks for MVI detection in hepatocellular carcinoma. Patches in WSI are first analyzed by pixel-level cancer tissue segmentation, followed by region-level instance feature extraction, and then by slide-level analysis using sc-GCN to solve the over-smoothing problem. Reviewers acknowedge the novelty of the cascade networks but have concerns about reproducbility (R1, R4 and AC), comparison to other methods (R2). In the rebuttal, I think authors have justified their existing comparison is already sufficient. Nevertheless, I share the same opinion with reviewers and meta-reviewer and consider it has a low reproducibility given neither the data nor the code is available. So it is difficult to judge how useful such a paper to audience given the complexity of the proposed system without code.
After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.

Reject
What is the rank of this paper among all your rebuttal papers? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers).

15

back to top

From Pixel to Whole Slide: Automatic Detection of Microvascular Invasion in Hepatocellular Carcinoma on Histopathological Image via Cascaded Networks