FAB-Net

FAB-Net represents segmentation networks combined with the foreground attention boosting (FAB) module. Here we provide the implementation code on COVID-19-CT-Seg dataset: 20 lung CT scans; Annotations include left lung, right lung and infections. And we evaluate our method on the infection segmentation task. For details, please refer to the segmentation task 1 of the website , including the 5-fold cross-validation, dataset split file, preprocessing method, and the evaluation metrics.

Task	Training and testing
Infection	5-fold cross validation 4 cases (20% for training) 16 cases (80% for testing)

For COVID-19-CT-Seg dataset, the baselines are based on the nnU-Net, which is a powerful segmentation method designed to deal with the dataset diversity found in the somain. It condenses and automates the keys decisions for designing a successful segmentation pipeline for any given dataset. In this repository, we apply the FAB module to nnU-Net without other special processing. One can refer to the websit for more implementation details of nnU-Net. For convenience, we provide one script demo.py for easy testing.

Installation

All the experiments are implemented via PyTorch with one Nvidia RTX 2080ti GPU, Ubuntu system and Python 3. One can refer to nnU-Net for more installation details or use the following process.

1. Install PyTorch (version 1.1.0)

2. Install Nvidia Apex by following instructions (after the installation of PyTorch):

   git clone https://github.com/NVIDIA/apex
   cd apex
   pip install -v --no-cache-dir ./

3. Install dependent packages and nnU-Net:

   git clone https://github.com/Womcos/FAB-Net.git
   cd FAB-Net
   pip install -r requirements.txt
   pip install -e .

Usage

The FAB-Net is based on the nnU-Net framework, and one can refer to the website for more implementation details.

Testing with pre-trained model

Here we provide the pre-trained model (password 9nhf )and the testing code of the FAB-Net on the COVID-19-CT-Seg dataset. For convenience, we provide a simple script demo.py for testing on the COVID-19-CT-Seg dataset. The parameters to be used in the script demo.py are as follows:

fold = '0'  ## '0','1','2','3','4'
test_folder = '.../COVID-19-CT-Seg/COVID-19-CT-Seg_20cases_edit1'

where test_folder is the folder of preprocessed raw images (10 cases from Coronacases are adjusted to lung window [-1250,250], and then normalized to [0, 255] ). The preprocessing code for raw images is given in the ./data_edit/demo.py. The parameter fold represents the fold number of the pre-trained model according to the dataset split method of the COVID-19-CT-Seg dataset. To use the pre-trained model for testing, one needs to set the base folder in ./nnunet/paths.py as the address of the pre-trained model folder .../nnunet2_COVID19_FAB and set the test_folder as the folder of images for testing. After running demo.py, the test results will be saved in the folder .../nnunet2_COVID19_FAB/test_results. Then switch the fold from '0' to '4' to get results of all folds. For convenience, we provide the testing results of the validation data for all folds in the folder .../validation_raw of the pre-trained model (password 9nhf ). One can easily evaluate the results without testing. In addition, the evaluation metrics are Dice similarity coefficient (DSC) and normalized surface Dice (NSD), and the python implementations are here which is given by the COVID-19-CT-Seg dataset.

Results

The quantitative results of nnU-Net baseline and FAB-Net for infection segmentation task are presented as follows:

Methods	Metrics	Fold 0	Fold 1	Fold 2	Fold 3	Fold 4	Avg
nnU-Net	DSC	68.08%	71.32%	66.18%	68.13%	62.67%	67.28%
	NSD	70.88%	71.82%	71.71%	70.84%	64.93%	70.04%
FAB-Net	DSC	73.59%	74.10%	73.67%	72.12%	66.30%	71.95%
	NSD	76.05%	74.47%	80.23%	75.42%	68.73%	74.98%

The quantitative results of nnU-Net baseline and FAB-Net for lung segmentation task are presented as follows:

Methods	Lung	Metrics	Fold 0	Fold 1	Fold 2	Fold 3	Fold 4	Avg
nnU-Net	Left	DSC	84.88%	80.28%	87.14%	88.44%	88.33%	85.82%
		NSD	68.69%	61.82%	74.34%	75.18%	75.83%	71.17%
	Right	DSC	85.21%	83.88%	90.34%	89.86%	90.22%	87.90%
		NSD	70.55%	68.25%	78.45%	78.45%	78.31%	74.80%
FAB-Net	Left	DSC	88.61%	86.50%	91.76%	88.79%	91.10%	89.35%
		NSD	75.26%	71.65%	79.50%	76.17%	78.19%	76.15%
	Right	DSC	89.55%	88.60%	92.51%	90.98%	91.79%	90.67%
		NSD	76.25%	74.87%	80.47%	78.49%	79.60%	77.94%

Questions

Please contact dxfeng@shu.edu.cn