Back to top List of papers List of papers - by topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Neda Zamanitajeddin, Mostafa Jahanifar, Nasir Rajpoot
Abstract
Digitization of histology images and the advent of new computational methods, like deep learning, have helped the automatic grading of colorectal adenocarcinoma cancer (CRA). Present automated CRA grading methods, however, usually use tiny image patches and thus fail to integrate the entire tissue micro-architecture for grading purposes. To tackle these challenges, we propose to use a statistical network analysis method to describe the complex structure of the tissue micro-environment by modelling nuclei and their connections as a network. We show that by analyzing only the interactions between the cells in a network, we can extract highly discriminative statistical features for CRA grading. Unlike other deep learning or convolutional graph-based approaches, our method is highly scalable (can be used for cell networks consist of millions of nodes), completely explainable, and computationally inexpensive. We create cell networks on a broad CRC histology image dataset, experiment with our method, and report state-of-the-art performance for the prediction of three-class CRA grading.
Link to paper
DOI: https://doi.org/10.1007/978-3-030-87237-3_28
SharedIt: https://rdcu.be/cymaj
Link to the code repository
https://github.com/mostafajahanifar/SNA_MICCAI21
Link to the dataset(s)
https://warwick.ac.uk/fac/cross_fac/tia/data/crc_grading/
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose the application of a statistical network analysis method to describe the structure of the tissue micro-environment by modelling nuclei and their connections as a network. The method is applied to colorectal adenocarcinoma classification.
- 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.
- originality of the approach (first to used SNA for automatic cancer grading)
- high scalability (can be used for cell networks consist of millions of nodes)
- explainability of the results
- computationally inexpensive
- comparison to many different other approaches
- 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 application to WSI
- 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
Code will be publicly available.
- 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 paper is very clear and the presented approach is really interesting as it is not based on deep learning architecture (difficult and expensive to train). But the most interesting aspect is the interpretability of the results, as they are based on features from social networks studies. It is then quite easy to analyse the results and obtain useful information for the pathologist. The only thing I would regret is that the method has not been applied to a WSI as many open dataset could be found.
- 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?
The paper is clear and the idea very original. The topic of analysing TME and interactions between cells is a topical subject. The interpretability of the approach is a real strength. Experiments are well described and could easily be reproduced.
- What is the ranking of this paper in your review stack?
1
- Number of papers in your stack
1
- Reviewer confidence
Very confident
Review #2
- Please describe the contribution of the paper
The authors propose to use a statistical network analysis method to describe the complex structure of the tissue micro-environment by modelling nuclei and their connections as a network. The paper shows that by analyzing only the interactions between the cells in a network, highly discriminative statistical features for CRA grading can be extrated. The authors create cell networks on a broad CRC histology image dataset, experiment with the proposed method, and report state-of-the-art performance for the prediction of three-class CRA grading.
- 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.
According to the authors, this is the first study that proposes the analysis of histology images using SNA measures for automatic CRA grading, demonstrating the role of cells and cellular communities as actors. Unlike other deep learning or convolutional graph-based approaches, the proposed method is highly scalable (can be used for cell networks consist of millions of nodes), completely explainable, and computationally inexpensive. Good result analysis (section 3.3)
- 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.
There is no homogeneous depth in the subsections of Section 2: there is so much depth in 2.3, while there is lack of details in 2.2.
- 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
Reproducibility of the proposed method is very limited, there are no enough details and many arbitrary decisions (or not explained). For example, there is clarity about the reason or justification of 180 features from SNA measures for each cell networks. A public dataset is used.
- 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
There are no enough details and many arbitrary decisions (or not explained). For example, there is clarity about the reason or justification of 180 features from SNA measures for each cell networks. Authors should review the depth of subsections 2.2 and 2.3 to incorporate more details in order to get reproducibility of the proposed method.
- Please state your overall opinion of the paper
Probably accept (7)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Very interesting approach, robust experimental framework, good result analysis.
- 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
Review #3
- Please describe the contribution of the paper
This paper presents a method histology image classification with measures from the social network analysis (SNA). All detected cells and spatial information are translated to a graph (i.e. cell social network). SNA measures and SNA-based statistical features are derived from each cell social network. The resulting features are used by SVM for histology image classification. The developed method is tested on colorectal adenocarcinoma cancer grade classification. Other baseline methods are compared with the proposed method variations where derived SNA features are aggregated and used in different ways. The resulting method presents good performance. Additionally, the resulting discriminating features and their distribution difference across different image classes are presented and analyzed.
- 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 work is aimed to characterize the tissue contextual information by social network analysis measures from a graph on cell-cell interactions.
- This paper focuses on feature interpretability and feature biology significance understanding.
- The proposed method is not subject to GPU memory limit.
- The performance of the proposed method with various ways of using and aggregating social network analysis measures are compared with other baseline 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.
- Although the construction of cell-cell connection graph does not require accurate cell segmentation, it still depends on an accurate cell detection. It is not clear how the end classification performance would be affected by erroneous cell detection results.
- As tumors are highly heterogeneous, there are multiple types of cells in presence of tissues. As a result, it is important to identify such cell population labels before such cell social network construction. However, this paper does not take this into account.
- Although it is a good idea to derive features from cell social network, such method may not be directly extended to other critical histology components that are large by size and limited by number.
- Although this paper presents good discussions on the biology significance of social network analysis features, it does not provide discussions on why one feature aggregation method would beat other method by classification performance.
- 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
The reproducibility of this paper is moderate, as certain key information is missing, e.g. the number of fold for cross-validation, training-testing data split, number of patients in the cohort, number of samples in each class, and selected features for classification.
- 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
- Although the construction of cell social network in section 2.2 does not require an accurate cell segmentation, it does need a precise cell detection. Thus, the cell detection performance can affect the end system classification performance. However, it is not clear to what extent an erroneous detection result would impact the classification accuracy.
- When the number of cells in an image is overwhelming large, the size of adjacency matrix A could be very large. As a result, such implementation technique as sparse matrix may be required for scalability. Some discussion on this implementation aspect would be necessary.
- There are multiple cell populations in tumor regions. Therefore, taking features from a graph with cells of unlabeled cell class may reduce the interpretability power of the resulting social network analysis features.
- The selected features and numbers of features for each aggregation method in section 2.5 are missing. It would be interesting to see if there would be some strong features selected in all these three feature sets.
- For the colorectal adenocarcinoma cancer data, such information as number of patients in each grade in this study, the number of fold for cross-validation, and training-testing data split is missing. Applying the developed method to a separate dataset for validation would be very helpful to show the generalizability of the proposed methods.
- It is not clear why the performance of one aggregation method is better than others in section 3.1 Some clarification or analysis would be highly useful.
- Only SVM is used in the experiment. Additional classifiers could be used to see if similar results can be achieved.
- This paper presents really nice discussions on feature biology significance in Section 3.3 However, the proposed method only strengthens interoperability of cell related features. It is not straightforward to directly extend this method to includes features of other histology primitives, especially those in large scale by size and limited by number.
- 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?
Although this paper is subject to certain weaknesses mentioned above, the way it derives features from histology component interactions is interesting. Because of its method design, it presents a strong ability to characterize the complex component spatial connections in tumor micro-environment and interpret the biological significance of the resulting discriminating features.
- What is the ranking of this paper in your review stack?
4
- Number of papers in your stack
5
- Reviewer confidence
Very confident
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.
Reviewers came to consensus on the excellent merits of this work. Please check the detailed comments by the reviewers and update the paper accordingly.
- 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).
4
Author Feedback
We thank the reviewers and the meta-reviewer for their time and effort in providing detailed commentary on our paper. We are pleased to see that the paper is provisionally accepted. We would like to take this opportunity to address a couple of points raised by the reviewers.
Reviewer 2 has raised a concern about the lack of justification for the number of SNA features. We computed a list of SNA features that we considered to be relevant and important. That, however, is by no means an exhaustive list of features. Having said that, it is the job of a good feature selection algorithm to detect the most relevant features and reduce the size of feature space considerably. The reviewer has also raised a concern about the lack of depth in Section 2.2. We decided not to add details about cell network construction in Section 2.2 since the construction is based on a previous work [14].
Reviewer 3 has complained about the lack of cross-validation details. We would like to remind the esteemed reviewer that those details are already provided at the end of Section 2.1 (Data). The reviewer has also asked for details about the dataset, experimental setting and a list of strong features. We will provide all those details in the camera-ready version. Finally, the reviewer has also made excellent suggestions about the comparison of different aggregation methods, object graphs, evaluation of robustness and scalability of the proposed method to inaccuracies in locations of detected nuclei. We appreciate these and would like to investigate these points in future. We thank the reviewers and the meta-reviewer for their time and effort in providing detailed commentary on our paper. We are pleased to see that the paper is provisionally accepted. We would like to take this opportunity to address a couple of points raised by the reviewers.
Reviewer 2 has raised a concern about the lack of justification for the number of SNA features. We computed a list of SNA features that we considered to be relevant and important. That, however, is by no means an exhaustive list of features. Having said that, it is the job of a good feature selection algorithm to detect the most relevant features and reduce the size of feature space considerably. The reviewer has also raised a concern about the lack of depth in Section 2.2. We decided not to add details about cell network construction in Section 2.2 since the construction is based on a previous work [14].
Reviewer 3 has complained about the lack of cross-validation details. We would like to remind the esteemed reviewer that those details are already provided at the end of Section 2.1 (Data). The reviewer has also asked for details about the dataset, experimental setting and a list of strong features. We will provide all those details in the camera-ready version. Finally, the reviewer has also made excellent suggestions about the comparison of different aggregation methods, object graphs, evaluation of robustness and scalability of the proposed method to inaccuracies in locations of detected nuclei. We appreciate these and would like to investigate these points in future.