Awesome

PF-Net: Pulmonary Fibrosis Segmentation Network with Multi-Scale Guided Dense Attention

This repository provides source code of PF-Net for pulmonary firbrosis segmentation proposed published on IEEE TMI 2021. If you use this code, please cite the following paper:

[1] G. Wang, S. Zhai et al. Semi-Supervised Segmentation of Radiation-Induced Pulmonary Fibrosis from Lung CT Scans with Multi-Scale Guided Dense Attention, IEEE Transactions on Medical Imaging, 2021.

pfnet_result

pf_net The structure of PF-Net. It combines 2D and 3D convolutions to deal with images with anisotropic resolution. For example, the in-plane resolution is around 4 times of through-plane resolution in our dataset, and we use 2D convolutions for the first two levels and 3D convolutions for the other levels in the encoder. Multi-Scale Guided Dense Attention is introduced in the decoder to deal with lesions with various positions, sizes and shapes.

Requirements

Pytorch version >=1.0.1.
PyMIC, a Pytorch-based toolkit for medical image computing. Version 0.2.4 is required. Install it by pip install PYMIC==0.2.4.
Some basic python packages such as Numpy, Pandas, SimpleITK.

Train and Test

Prepare your dataset and write .csv files for training, validation and testing. See config/data_train.csv for example.
Edit config/pfnet.csv, set the data root and csv files according to your computer. You may also need to set train_transform and test_transform based on the preprocess strategies required by your dataset.
Run the following commands for training and inference:

python net_run.py train config/pfnet.cfg
python net_run.py test config/pfnet.cfg