motikadosh / mtl-segthor

Dependent Multi-Task Learning for the Segmentation of Thoracic Organs at Risk in CT Images

MTL-SegTHOR

Dependent Multi-Task Learning for the Segmentation of Thoracic Organs at Risk in CT Images

author: Tao He

Institution: Sichuan University

email: [email protected]

Tookit need Python 3, pytorch 1.1.0

Prepare SegTHOR dataset

reading preprocessing.ipynb in details

the source data download from https://ent.normandie-univ.fr/filex/get?k=oZgYIeT5lnbxhtHZ2u8
or download from my Baidu Netdisk https://pan.baidu.com/s/1dQHYKIkUd5qCXIvdxSijNg; password: i41q
The data path organized like:
../data/data_source/Patient_01/GT.nii
../data/data_source/Patient_01/Patient_01.nii
Then, using preprocessing.ipynb to process the SegTHOR dataset for 4-fold cross-validation

Network training

See main.py parser.argument in details

A test run is like this
python3 main.py -b 16 --gpu 0,1,2,3 --model_name ResUNet101 --save_dir SavePath/SM --lr 0.01 --if_dependent 0 --if_closs 0

Important parameters:

• model_name -> indicate which encoder network is used. Please check models/model_loader.py in details.
• if_closs -> if using the classification loss (MTL). if_closs=0, MTL is not deployed.
• if_dependent -> if using the WMCE loss function. if_dependent=0, binary relevance with BCE loss functions is used.

Network testing

See prediction.ipynb in details

Please don't hesitate to contact me if you have any question about the data, method, or code !