Self-supervised nuclei segmentation

Code to train a self-supervised segmentation network for segmentation of nuclei in histopathology images [1].

• train.py contains training code and defines command line options.
• datasets.py defines datasets used to read images.
• models.py defines relevant models (attention network and scale network).
• utils.py defines extra useful functions.
• configs/ defines .yaml configuration files to set experiment parameters.

Installation

The Anaconda environment is specified in conda_env.yml. The environment can be recreated using

conda env create -f conda_env.yml

Tested with Nvidia GeForce GTX 1080 and GeForce GTX 1080 Ti GPUs, running driver version 410.48 and cuda 10.0, and Pytorch 1.1.0 with torchvision 0.3.0.

Data

Please see the directory data processing for instructions on downloading and using data.

Usage

train.py is the training code which offers three command line parameters.

• --cfg specifies the configuration file to use.
• --gpu specifies which GPU to use. A value of -1 implies no GPU.
• --output_dir specifies directory to record results. If the configuration file is name.yaml, results will be recorded in <output_dir>/name.

Example train usage---

python train.py --cfg configs/example.yaml --gpu 0 --output_dir /path/to/output/

Example testing usage---

python test.py --cfg configs/example.yaml --epoch 200 --dataroot /path/to/test/imgs/ --ext tif --gpu 0 --output_dir /path/to/output/

Config files

See configuration options for a description of configuration options

Reference

[1] Self-Supervised Nuclei Segmentation in Histopathological Images Using Attention, Mihir Sahasrabudhe, Stergios Christodoulidis, Roberto Salgado, Stefan Michiels, Sherene Loi, Fabrice Andre, Nikos Paragios, Maria Vakalopoulou, MICCAI 2020 [ PDF ]