This is the code for MICCAI Automatic Prostate Gleason Grading Challenge 2019. Check here and here, we took the 1st place of task 1: pixel-level Gleason grade prediction and task 2: core-level Gleason score prediction (leaderboard).
Task 1 is regarded as a segmentation task, and we use PSPNet for this. And for task 2, we do not train a different network, but just produce the prediction from the prediction of task 1 according to the Gleason grading system.
The train script is based on reference script from torchvision 0.4.0 with minor modification. So, you need to install the latest PyTorch and torchvision >= 0.4.0. Check requirements.txt for all packages you need.
This repo use GluonCV-Torch, thanks for Hang Zhang's outstanding work!
Each image is annotated in detail by several expert pathologists. So how to use this annotations is important. We use STAPLE to create final annotations used in training. Check the preprocessing.py script for detail.
To run the training, simply run python train.py, check python train_gleason.py --help for available args.
To run the inference, simply run python inference.py, check python inference.py --help for available args.