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

Authors

Pingjun Chen, Muhammad Aminu, Siba El Hussein, Joseph Khoury, Jia Wu

Abstract

The cells and their spatial patterns in the tumor microenvironment (TME) play a key role in tumor evolution, and yet remains an understudied topic in computational pathology. This study, to the best of our knowledge, is among the first to hybrid local and global graph methods to profile orchestration and interaction of cellular components. To address the challenge in hematolymphoid cancers where the cell classes in TME are unclear, we first implemented cell level unsupervised learning and identified two new cell subtypes. Local cell graphs or supercells were built for each image by considering the individual cell’s geospatial location and classes. Then, we applied supercell level clustering and identified two new cell communities. In the end, we built global graphs to abstract spatial interaction patterns and extract features for disease diagnosis. We evaluate the proposed algorithm on H&E slides of 60 hematolymphoid neoplasm patients and further compared it with three cell level graph-based algorithms, including the global cell graph, cluster cell graph, and FLocK. The proposed algorithm achieves a mean diagnosis accuracy of 0.703 with the repeated 5-fold cross-validation scheme. In conclusion, our algorithm shows superior performance over the existing methods and can be potentially applied to other cancer types.

Link to paper

DOI: https://doi.org/10.1007/978-3-030-87237-3_16

SharedIt: https://rdcu.be/cyl9V

Link to the code repository

https://github.com/WuLabMDA/HierarchicalGraphModeling

Link to the dataset(s)

N/A

Reviews

Review #1

Please describe the contribution of the paper

Based on digital pathological images, this paper proposes a hierarchical graph construction algorithm that characterizes cellular architectures at local cellular level and global community level, and the features extracted from this graph show superior diagnostic performance of lymphoid neoplasms compared with other graph theory based methods.
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 demonstates the feasibility of cell-level and supercell-level graph based features for diagnosing lymphoid neoplasms.
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 novelty of the method is not enough. Similar ideas of using unsupervised clustering to discover cell subtypes and using occurrence counts of different types of edges in Delaunay triangulation graph as features have been explored in prior work (PMID: 29136101).
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

This paper provides sufficient details about the method, dataset, and evaluation.
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
1. In table 1, the authors say the number of cell features is 24, but the number of listed feature names seems to be only 23. Please check.
2. In the steps of cell and supercell phenotyping, the performance of classifiers should be reported.
3. What were the criteria in selecting the two parameters used in the ROC algorithm?
4. The authors identified two intrinsic cell phenotypes. They should also discuss with a pathologist if there are any biological meanings related to the two cell phenotypes.
5. In addition to comparing the proposed method with graph-based methods, I suggest the authors include some simpler features as baselines that may also be quite effective. For example, the authors can use the L1-normalized histograms of each of the 24 cell features as image-level features and train classifiers on the image-level features.
Please state your overall opinion of the paper

borderline reject (5)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The method is not of sufficeint novelty and the (intermediate) results lack biological interpretations.
What is the ranking of this paper in your review stack?

4
Number of papers in your stack

5
Reviewer confidence

Very confident

Review #2

Please describe the contribution of the paper

This study leverages a hybrid local and global graph-based representation for the study of hematolymphoid cancers from H&E images. The paper addresses a challenging problem for a particular type of cancer where cell phenotyping and TME characterization is not well established. The paper is well-written, well-structured, and easy to follow. The overall design is complex and well-motivated by the challenges induced by local TME through ITH, thus proposing to characterize cell patterns interaction at multiple levels (introducing the supercell concept) with hierarchical graph-based modelling and unsupervised clustering for cell community detection.
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 study brings forward a novel methodology for cell, clonal level phenotyping and cell-interaction patterns, based on a complex design including unsupervised clustering and hierarchical graph-based modelling. While the underlying approaches are not entirely new, the combination of those to form a pipeline that leverages multilevel feature extraction for cell profiling and cell communities pattern interaction characterization is very solid. This is particularly useful for cancers where cell profiles are unclear and TME is less studied, such as hematolymphoid neoplasms, the authors being the first to address its challenges through the proposed method. Performance comparisons with recent graph-based cell clustering approaches were provided, the proposed method achieving superior performance.
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 overall methodology design seems quite powerful, however, the choice of certain approaches and/or specific parameters is sometimes unclear. Although the aim of the paper is to address a very specific application with particular challenges, hematolymphoid neoplasms, it is not clear whether the present approach would be suitable to study different types of cancer with a better studied TME and known cellular profile, the main concern being thus a limited scope. The authors mention previous attempts based on GCN approaches for automatic feature extraction at multiple levels from cellular level graphs. Known to be powerful methods, capable of inferring different features at different levels, a comparison to at least one of these methods would have been interesting to better showcase the potential of their proposed techniques.
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

Authors specify that the data and code will be made publicly available for reproducing 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

As previously mentioned, the main concern is the applicability to the present framework for the study of different cancers, where cellular profiling is better studied. Perhaps more clarity for the choice of certain approaches and comparison to one GCN approach after cell type detection, would have helped build a stronger case. In particular, more clarity would have been expected in a few cases to justify the use and adaptability to various contexts: e.g. supercell via local graph, it is not clear what cellular feature vector represents in this context and why it is defined as belonging to R3 . In section 2.3, more details could have facilitated the understanding of otherwise interesting choices: the cellular spatial orchestration and interaction with Voronoi diagrams, it is not clear how they are constructed (are local cell localisations within a supercell the seeds for Voronoi diagram?). Details on the intrinsic parameters of t-SNE, and why that is preferred to similar methods would have helped clarify the results, as well as the sensitivity of the results w.r.t to the parameter choice, i.e. 2 types of cell communities.
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?

The authors introduced a novel, solid hybrid unsupervised framework for cell phenotyping and cellular community detection, addressing the challenges of hematolymphoid cancers and demonstrating its superior performance in comparison to recent approaches. Providing supplementary justification for the use of selected methods together with the configuration of intrinsic parameters would build a stronger case for adaptability to other cancer types. Comparison to more complex methods (e.g. GCN) would strengthen the method proposal.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

6
Reviewer confidence

Very confident

Review #3

Please describe the contribution of the paper

The authors propose a hierarchical modelling of cells and cell interactions to understand the tutor micro environment and intratumoral heterogeneity of large B cell lymphoma. Their method is based on clustering to recognise the intrinsic types of cells, and building cell communities (supercells, that are also clustered) on which they compute a graph of supercell interactions. The graph representation is used for image classification using an SVM.
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 approach is technically sound and very interesting.
- The problem is clinically significant and the description of the relevant literature is quite complete.
- The choice of methods, although not novel, makes sense as a whole, and allows understanding what the purpose is.
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.
- It is not clear if the phenotyping was also cross-validated, which would have made the method generalises well to unseen data.
- Understanding the space limitations, I miss a more in depth discussion of the results when compared to the other methods. It is very surprising to see almost 0.4 gain on the CLL class and not having an hypothesis about why.
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

Since the methods are based on clustering, it would be difficult to reproduce exactly the same results. By reporting the mean and standard deviation of the results after cross validation, it is expected that a fair comparison can be done by other researchers. Code and data are expected to be released after publication.
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

I liked the paper overall, and the strengths compensate the weaknesses (see above). One thing I would recommend is to have a more in depth discussion on which methods are selected to compare and why.
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 makes a significant contribution. Even if using standard tools, the architecture is novel.
What is the ranking of this paper in your review stack?

2
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.

There are mixed opinions. One reviewer noted lack of novelty. There are also some concerns regarding performance comparison and result discussion from first and third reviewers. These issues should be addressed in the rebuttal.
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).

2

Author Feedback

We appreciate the encouraging comments from reviewer 2 and reviewer 3, noting that the key innovation of our study in hierarchical modeling at the local (cell-cell interaction) and global (cell community interaction) levels to characterize tumor composition, microenvironment, and intratumoral heterogeneity with a diagnostic application in a prevalent type of lymphoid neoplasm. We notice that the major concern raised by reviewer 1 is due to our miscommunication, which we will provide detailed explanations below.

Reviewer 1: Lack of novelty, similar ideas have been explored in prior work.

We’d like to thank reviewer 1 for bringing up the prior work, which gives us a chance to highlight the novelty of our study. There are several innovations in our study which distinguish our work from the prior work. In the study by Cheng et al [1], the authors built a global graph (Delaunay) to quantify only local cell-cell interactions and used these patterns for clinical evaluation, which is in line with the Flock algorithm [2] that shows inferior performance compared to our algorithm. By contrast, we propose a hierarchical model to comprehensively characterize interaction patterns at both the local cellular level and global community level for cellular architectures. To the best of our knowledge, this is the first study using such a design in the digital hematopathology field. Further, we have proved that by combining both local and global information, we can significantly improve the model’s performance. Therefore, we believe that our design has opened a new direction to quantify intratumoral heterogeneity, an aspect that has particular clinical implications in terms of risk stratification and prediction. Indeed, these are the specific contributions of our work that were recognized by both of the other reviewers, where they have commented that the proposed architecture of hierarchical phenotyping and graph modeling to characterize the spatial architecture in lymphoid neoplasms is novel and solid. [1] Cheng, J. et al., Bioinformatics, 34.6 (2018): 1024-1030. [2] Lu, C. et al., Medical Image Analysis 68 (2021): 101903.

Reviewer 1: Besides comparison with graph-based methods, the authors shall compare simpler features as baselines. Reviewer 3: An in-depth discussion of the results compared to the other methods would be beneficial. What’s the hypothesis on almost 0.4 gain on the CLL disease?

We follow reviewer 1’s suggestion, use the simple features as baseline (24 cell features + kernel SVM), and obtain the mean acc of 0.483, CLL AUC of 0.712, aCLL AUC of 0.556, and RT DLBL of 0.722 following an identical testing scheme. This observation suggests that the individual cell features alone without consideration of spatial interactions are not sufficient to characterize a complex tumor microenvironment. As nicely pointed out by reviewer 3, using only the local or the global graph failed to separate CLL from the other clinically relevant subtypes built into the design of our study. By contrast, combining the local and global graph can improve the performance by a significant margin with a 0.4 increase in accuracy on the CLL. We have conducted further investigations and found that the newly identified two cell communities have significantly different distributions between CLL and aCLL + RT (p-value = 3.54e-08). Hence, the community-derived features can accurately separate CLL from other clinically relevant subtypes. Hereby, we hypothesize that the proposed hybrid design can overcome the limitation inherent in the adoption of the global or local graph approaches solely to more meaningfully profile tumor composition and intratumoral heterogeneity. Indeed, the global approach misses the cellular level details while the local graph ignores the high-level interaction patterns between communities.

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.

The rebuttal has provided good responses regarding novelty and evaluation. The overall paper quality is good.
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).

1

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.

The novelty concern seems to be addressed well. I do agree with R2 that comparison with GCN would strengthen the paper. Overall the paper should be accepted.
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).

10

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.

Most of reviewers acknowledge the contribution of the papers and authors have also further address the issue of R1 and stress the novelty of the manuscript as compared to the state-of-the-art. I think the research problem is very important and also appreciate the soundness of the paper/rebuttal and support its acceptance.
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).

2

back to top

Hierarchical Phenotyping and Graph Modeling of Spatial Architecture in Lymphoid Neoplasms