Colonoscopy tissue screen and segmentation. [DigestPath 2019] [2nd place method]

This respository is PyTorch implementation of our method in Task 2: Colonoscopy tissue segmentation and classification of DigestPath 2019. We won 2nd place in the competition,

Requirements

• Python3 (Python3.7 is recommended)
• PyTorch >= 1.1 PyTorch1.1, Python3.7

Installation:

• create conda environment: conda create -n torch1.1 python=3.7
• activate conda environment: source activate torch1.1
• install PyTorch1.1 following https://pytorch.org/get-started/locally/
• install other package pip install -r requirement.txt

Usage

Prepare data

1. Download data: Follow the instruction here.

2. Prepare data: unzip the download file, and put all the image together in data/raw_data.

   note: only malignant tissue have corresponding mask, you need to generate mask for benign tissue (zero mask with the same size) to train malignant and benign sample together. You can use use script lib/util/process/pre_processes/generate_benign_mask.py

Training model

Quick start

CUDA_VISIBLE_DEVICES=0 python train_net.py --config-file ./configs/unet/layer4_dicelossv1_dsv2_down4.yaml

This will train our final submitted model on fold 0 (total 4 fold). It will take about 5GB GPU memory for this configuration.

Model, training and validation result will be saved in tensorboard in cfg.OUTPUT_DIR, use tensorboard to have a look.

configure your own model

To use other model, se configs/defaults.py for default configuration. You can create your own yaml file to overwrite default configuration. Eg. set cfg.MODEL.MODEL = 'UNet16Layer4' to use our proposed model

Test

Use the same yaml file to test the trained model.

eg. CUDA_VISIBLE_DEVICES=0 python test_net.py --config-file ./configs/unet/layer4_dicelossv1_dsv2_down4.yaml This will give dice score and auc of the trained model on test dataset.

set cfg.SOLVER.DRAW to True to save visualization result as image in cfg.OUTPUT_DIR for a better view.