Authors

Yawen Wu, Yingli Zuo, Qi Zhu, Jianpeng Sheng, Daoqiang Zhang, Wei Shao

Abstract

Whole-Slide Histopathology Image (WSI) is regarded as the gold standard for survival prediction of Breast Cancer (BC) across different subtypes. However, in cancer prognosis applications, the cost of acquiring patients’ survival information is high and can be extremely difficult in practice. By considering that there exists certain common mechanism for tumor progression among different subtypes of the Breast Invasive Carcinoma(BRCA), it becomes critical to utilize data from a related subtype of BRCA to help predict the patients’ survival on target domain. To address this issue, we proposed a TILs-Tumor interactions guided unsupervised domain adaptation (T2UDA) algorithm to predict the patients’ survival on the target BC subtype. Different from the existing feature-level or instance-level transfer learning strategy, our study considered the fact that the tumor-infiltrating lymphocytes (TILs) and its correlation with tumors reveal similar role in the prognosis of different BRCA subtypes. More specifically, T2UDA firstly employed the Graph Attention Network (GAT) to learn the node embeddings and the spatial interactions between tumor and TILs patches in WSI. Then, besides aligning the embeddings of different types of nodes across the source and target domains, we proposed a novel Tumor-TILs interaction alignment (TTIA) module to ensure that the distribution of interaction weights are similar in both domains. We evaluated the performance of our method on the BRCA cohort derived from the Cancer Genome Atlas (TCGA), and the experimental results indicated that T2UDA outperformed other domain adaption methods for predicting patients’ clinical outcome.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43987-2_59

SharedIt: https://rdcu.be/dnwKe

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #2

Please describe the contribution of the paper

This paper proposes an unsupervised domain adaptive algorithm for survival analysis of patients with different subtypes of BRCA. The proposed method is based on the construction of WSI-level graphs. During model training, the error between source and target subtype is considered, and the source subtype loss function is modified to improve the survival prediction performance of the model from source subtype to target subtype. Experiments on BRCA datasets from The Cancer Genome Atlas (TCGA) demonstrate the effectiveness of the 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 idea of utilizing common mechanisms between different subtypes of BRCA to achieve cross-subtype adaptive transfer learning looks novel and interesting in pathology image analysis.
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) Explanations are not clear and solid. For example, in data preprecessing, it not clear how to segment TILs and tumor tissues by the U-Net++ model. 300 patches are selected. Top 10 percent connections are selected. No ablation study had been performed to justiy these pre-defined key parameters. (2) Comparions had not performed by using state-of-the-are survival prediction models. As shown in Table 1, all compared methods had been proposed several years ago.
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 reproducibility is not so well, considering that their codes are not public accesible and the method description is not very clear.
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

(1) More comparions with state-of-the-are survival prediction models should be provided. (2) How to train the UNet++ to segment the TILs and Tumor tissues? How to determine those pre-defined parameters used in data pre-processing? (3) In the abstract, the first sentence states that ‘WSI is regarded as the gold standard for survival prediction of BC’. To my opinion, WSI is the gold starndared for cancer diagnosis, but maybe it is not the glod starndared in survival predcition. (4) Some symbols should be corrected. For example, in the page 4, the symbol Ni in the sentence following the equation (2) should be made consistent with that in equation (2).
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?

experimental comparisons and methological descriptions
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

N/A
[Post rebuttal] Please justify your decision

N/A

Review #1

Please describe the contribution of the paper

The papers propose a new methodology to perform domain adaptation. In particular, it is related to the research field of transfer learning to tackle the lack of annotated data in histopathology tissue 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.

The theoretical part of the paper is excellent and uses graph representations of interactions between TILs and Tumor cells. Taking into account the spatial context within the micro-tumoral environment is a fundamental aspect of modern digital histopathology.

The results are superior to SOTA AUC and specially relatively to feature-alignment methods, reaching 0.7 on a difficult problem (based on HE slides): one main explanation of this is the minimization of differences in TILs-Tumor interaction weights.
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 conclusion stating that “To the best of our knowledge, T2UDA was the first method to successfully achieve interrelationship transfer between TILs and tumors across different cancer subtypes for prognosis task” is perhaps a bit overstated. For instance : (Parreno-Centeno et al. 2022) A deep learning and graph-based approach to characterize the immunological landscape and spatial architecture of colon cancer tissue Mario Parreno-Centeno, Guidantonio Malagoli Tagliazucchi, Eloise Withnell, Shi Pan, Maria Secrier bioRxiv 2022.07.06.498984; doi: https://doi.org/10.1101/2022.07.06.498984

(Martin-Gonzalez et al. 2021) Predictive modelling of highly multiplexed tumour tissue images by graph neural networks, Paula Martin-Gonzalez, Mireia Crispin-Ortuzar, Florian Markowetz medRxiv 2021.07.28.21261179; doi: https://doi.org/10.1101/2021.07.28.21261179

etc.
Please rate the clarity and organization of this paper

Excellent
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
I would add a description of the subtypes of BRCA earlier in the paper to understand more easily the practical aspect of the research. A few questions to make the paper clearer :
- In the data preprocessing, 300 patches were chosen regardless of the number of patches. It would be nice if the authors could explain why 300 was chosen rather than some other fixed number or percentage of the total number of patches.
- It is not quite clear which survival was used as a target, overall survival or progression fee survival or other?
- Explain a bit more how the imbalance tbetween ER+ and ER- is tackled.
- In Data preprocessing, it is mentionned that the top 10% connections with the smallest distance, but which metric has been chosen.
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

7
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The interaction modeling with graph interaction for transfer learning is excellent and the score are superior to SOTA for the practical clinical test.
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 #3

Please describe the contribution of the paper

This paper takes advantage of the common histological manifestations and mechanisms between different subtypes of invasive breast cancer to predict patient survival status, which is challenging to collect information on. In this practical case, the paper validates the feasibility of graph structure features being transferable across different pathological subtypes.
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 paper validates the interaction of pathological image features in graph structures in a real-world task scenario with practical value, and this interaction is transferable between different cancer subtypes, which has reference significance for more similar scenarios.

The paper is well-organized and well-written, with interesting results presented in the experimental section.
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.

Similarly, the reported application scenario is too limited, and the adaptability of the proposed method to more generalized applications remains speculative.

The paper presents a practical application and points to interesting potential extensions, but the innovation is mediocre.

A minor weakness: The performance of the proposed method is heavily dependent on the TILs-Tumor Interaction Alignment module, but the paper does not adequately discuss what specific knowledge is being transferred in the TILs-Tumor Interaction Alignment module.
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 paper does not provide source code or state that source code will be provided. I believe, the reproducibility of the proposed method is medium without open-source code.
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 follow these questions:
1. In the TILs-Tumor Interaction Alignment module, can more detailed explanations be given for the physical meanings of p_i and q_i?
2. The proposed method uses GAT to model the interaction features between tissues. GAT can reflect the importance of nodes and highlight essential patches. Does the paper discuss whether more graph topology information or important node feature information is transferred during the transferring process? I think a good answer to the first question can solve this problem as well.
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?

Overall, the paper presents a trustable method with a valuable practical application, slightly above the baseline.
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

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.

This work proposes an unsupervised domain adaptive algorithm, utilizing WSI-level graphs, to tackle the lack of annotated data and to conduct survival analysis of patients with different subtypes of BRCA. The paper is, in general, well written and presented. The idea of utilizing common mechanisms between different subtypes of BRCA to achieve cross-subtype adaptive transfer learning is interesting. The authors also present the the interaction of pathological image features in graph structures in a real-world task scenario, which can be transferable between different cancer subtypes. However, there are a few concerns on the paper. The authors may rephrase the main texts since some of the statements are overstated, in particular in the conclusions. The authors may improve the description of the datasets, data processing, and method for better transparency and reproducibility of the work. The proposed method is heavily dependent on the TILs-Tumor Interaction Alignment module, but the paper does not adequately discuss what specific knowledge is being transferred in the TILs-Tumor Interaction Alignment module. The authors may provide their insights into this.

Author Feedback

N/A

back to top

Transfer Learning-Assisted Survival Analysis of Breast Cancer Relying on the Spatial Interaction Between Tumor-Infiltrating Lymphocytes and Tumors