GraphLSurv

Source code of the paper titled "GraphLSurv: A Scalable Survival Prediction Network with Adaptive and Sparse Structure Learning for Histopathological Whole-Slide Images" and published in Computer Methods and Programs in Medicine. The paper is available at DOI: 10.1016/j.cmpb.2023.107433.

TLDR

This work presented a scalable framework for estimating survival from patient's histopathological Whole-Slide Images (WSIs). In this framework, an adaptive and sparse graph structure is employed to describe each gigapixel WSI and then learn a global representation for prediction. This work absorbed an effecient anchor-based technique to handle very large-scale graphs. We proved that an adaptive and sparse graph structure could be more suitable for WSI-based survival prediction, compared to those densely-connected strutures like that adopted in Transformer.

GraphLSurv overview:

There are two steps for WSI-based survival prediction: WSI Preprocessing and GraphLSurv Modeling. We show them below.

WSI Preprocessing

In this setp, WSIs are converted into multiple image patches and instance features at first. This repo resolves this into four subtasks:

• WSI dataset exploring
• Tissue region segmentating and tiling
• Patch feature extracting
• Overview of the preprocessing result

There are three directories in S01-preprocessing:

• tools: We fork CLAM and have made it adapted to our task. The current version implements all preprocessing procedures, such as tissue region patching, patch energy calculation, sampling & feature extraction.
• scripts: It provides some essential command-lines for all preprocessing procedures.
• notebooks: Notebooks give the details of each preprocessing part, including their results, for better understanding how to preprocess WSIs.

GraphLSurv Modeling

In this step, the data from the first step will be input to the graph-based network for survival prediction. More details of this step could be found in GraphLSurv paper.

After you finished WSI Preprocessing (please ensure that you have followed the given instructions), you can train GraphLSurv by using the following commands (taking NLST dataset for example):

# build the initial KNN graph for each patient from NLST 
cd ./scripts 
## it is optional
./S1-Build-Graph.sh 

# train GraphSurv model on NLST
cd ..
## please set `age_ratio_init_graph` to 0 in a YAML file if you ignored the optional graph building.
python3 main.py --config config/report-nlst.yml --multi_run 

About data splitting: We provided the details of data splitting (all patient_id of training/val/test sets) used in experiments. Please see them in S02-modeling/data_split. As most methods proposed previously did not make their splitting of training/val/test publicly-available, it is really hard for one who wants to reproduce the reported results. For this reason, we released the data splitting used in our experiments.

Citation

Please cite this work if you used it in your research via

@article{LIU2023107433,
    author = {Liu, Pei and Ji, Luping and Ye, Feng and Fu, Bo},
    doi = {https://doi.org/10.1016/j.cmpb.2023.107433},
    issn = {0169-2607},
    journal = {Computer Methods and Programs in Biomedicine},
    pages = {107433},
    title = {{GraphLSurv: A scalable survival prediction network with adaptive and sparse structure learning for histopathological whole-slide images}},
    url = {https://www.sciencedirect.com/science/article/pii/S0169260723001001},
    volume = {231},
    year = {2023}
}

or P. Liu, L. Ji, F. Ye, and B. Fu, “GraphLSurv: A scalable survival prediction network with adaptive and sparse structure learning for histopathological whole-slide images,” Comput. Methods Programs Biomed., vol. 231, p. 107433, 2023, doi: https://doi.org/10.1016/j.cmpb.2023.107433.