HCCSurvNet: Deep learning predicts postsurgical recurrence of hepatocellular carcinoma from digital histopathologic images

This repository contains the code to quantify risk scores for recurrence in patients with hepatocellular carcinoma from H&E-stained FFPE histopathology images.

Software Requirements

This code was developed and tested in the following settings.

OS

• Ubuntu 18.04

GPU

• Nvidia GeForce RTX 2080 Ti

Dependencies

• python (3.6.10)
• numpy: (1.18.1)
• pandas (0.25.3)
• pillow (7.0.0)
• scikit-learn (0.21.3)
• scikit-image (0.15.0)
• scikit-survival (0.11)
• opencv-python (4.1.2.30)
• openslide-python (1.1.1)
• staintools (2.1.2)
• h5py (2.9.0)
• pytables (3.5.1)
• pytorch (1.4.0)
• torchvision (0.5.0)

Installation

Install Miniconda on your machine (download the distribution that comes with python3).

After setting up Miniconda, install OpenSlide (3.4.1):

apt-get install openslide-tools

Create a conda environment with environment.yml:

conda env create -f environment.yml

Activate the environment:

conda activate hccsurvnet

Demo

data collection

Download diagnostic whole-slide images from TCGA-LIHC project using GDC Data Transfer Tool Client.

gdc-client download -m gdc_manifest_tcga_lihc.txt

Download TCGA-CDR-SupplementalTableS1.xlsx from Integrated TCGA Pan-Cancer Clinical Data Resource and rename it to metadata.csv.

Get tumor region annotations on whole-slide images using Aperio ImageScope in XML format.

prepare dataset for tumor tile classification

python xml2tile.py  
python xml_tile2hdf.py  

train and evaluate tumor tile classifier

python tumor_tile_classifier.py

*** Output: AUROC (areas under the receiver operating characteristic curve) and its 95% confidence interval

prepare dataset (TCGA-HCC) for risk score prediction

python svs2tile.py  
python svs_tile2hdf.py  

apply tumor tile classifier to the TCGA-HCC dataset

python tumor_tile_inference.py  

select top-X tiles (Default: X=100)

python select_topX.py  

train and evaluate risk score predictor

python risk_score_predictor.py  

*** Output: Harrell's and Uno's C-indices

Note: please edit paths in each .py file.

License

This code is made available under the MIT License.

Citation

Scientific Reports 2021;11(1):2047

@ARTICLE{Yamashita2021deep,
  title = "Deep learning predicts postsurgical recurrence of hepatocellular carcinoma from digital histopathologic images",
  author = "Yamashita, Rikiya and Long, Jin and Saleem, Atif and Rubin, Daniel L and Shen, Jeanne",
  journal  = "Sci. Rep.",
  volume   =  11,
  number   =  1,
  pages    = "2047",
  month    =  jan,
  year     =  2021,
  language = "en"
  }