SegGini:

SEGmentation using Graphs with Inexact aNd Incomplete labels

This repository contains the code to reproduce the results of "Learning Whole-Slide Segmentation from Inexact and Incomplete Labels using Tissue Graphs", MICCAI, 2021.

The code is built on the Histocartography library, a python-based package that facilitates the modelling and learning of pathology images as graphs.

The described experiments are presented for the SICAPv2 dataset, a cohort of Hematoxylin and Eosin (H&E) stained prostate needle biopsies.

Overview

Installation

Cloning and handling dependencies

Clone the repo:

git clone [email protected]:histocartography/seg-gini.git && cd seg-gini

Create a conda environment and activate it:

conda env create -f environment.yml
conda activate seggini

Install PyTorch:

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

Downloading and preparing the SICAPv2 dataset

SICAPv2 is a database of H&&E stained patches (512x512 pixels) from 155 prostate whole-slide images (WSIs) across 95 patients. The dataset contains local patch-level segmentation masks for Gleason patterns (Non-cancerous, Grade3, Grade4, Grade5) and global Gleason scores (Primary + Secondary).

The stitched WSIs and corresponding Gleason pattern segmentation masks can be downloaded from SICAPv2. Alternatively, the WSIs and masks can be constructed, i.e., patch downloading and stitching, by running:

python bin/create_sicap_data.py --base_path <PATH-TO-STORE-DATASET>

A sample WSI and corresponding segmentation mask is demonstrated as follows. To highlight, the available Gleason score is inexact as it only states the worst and the second worst Gleason pattern present in the WSI.

Running the code

SegGini aims to leverage the WSI-level inexact supervision and incomplete pixel-level annotations for semantically segmenting the Gleason patterns in the WSI. To this end, first it translates a WSI into a Tissue-graph representation, and then employs Graph Neural Network based Graph-head and Node-head.

Step 1: Tissue-graph representation

The WSI to Tissue-graph transformation can be generated by running:

python bin/preprocess.py --base_path <PATH-TO-STORED-DATASET>

The script creates three directories with the following content per WSI:

• a tissue graph as a .bin file
• a superpixel map as a .h5 file
• a tissue mask as a .png file

Here, we also parse the available image and pixel annotations to create the necessary pickle files, to be used during training phase.

Finally, the directories should look like:

SICAPv2-data
|
|__ preprocess
|   |
|   |__ graphs
|   |
|   |__ superpixels 
|   |
|   |__ tissue_masks 
|
|__ pickles
|   |
|   |_ images.pickle
|   |
|   |_ annotation_masks_100.pickle 
|   |
|   |_ image_level_annotations.pickle
|
|__ images

Step 2: Training + Testing SegGini

We provide the option to train SegGini for three types of annotations.

1. Inexactimage-level annotations
2. Incompletepixel-level annotations
3. InexactandIncompleteannotations

Training SegGini for Inexact annotations:

python bin/train_graph.py --base_path <PATH-TO-STORED-DATASET> --config <PATH-TO-TRAINING-CONFIGURATION-YAML-FILE> 

Training SegGini for Incomplete annotations:

python bin/train_node.py --base_path <PATH-TO-STORED-DATASET> --config <PATH-TO-TRAINING-CONFIGURATION-YAML-FILE> 

Training SegGini for Inexact + Incomplete annotations:

python bin/train_combined.py --base_path <PATH-TO-STORED-DATASET> --config <PATH-TO-TRAINING-CONFIGURATION-YAML-FILE> 

Sample configuration yaml files for all the above cases are provided in ./config

Cite

If you use this code, please consider citing our work:

@inproceedings{anklin2021,
    title = "Learning Whole-Slide Segmentation from Inexact and Incomplete Labels using Tissue Graphs",
    author = "Valentin Anklin, Pushpak Pati, Guillaume Jaume, Behzad Bozorgtabar, Antonio Foncubierta-Rodriguez, Jean-Philippe Thiran, Mathilde Sibony, Maria Gabrani and Orcun Goksel",
    booktitle = "Medical Image Computing and Computer-Assisted Intervention (MICCAI)",
    pages = "636-646",
    year = "2021"
}