segmentor

Segmentor is a deep learning segmentation engine.

Install

# activate conda
conda activate
# clone the repository
git clone https://github.com/inqlee0704/segmentor.git

# Install packages
cd segmentor
python -m pip install -r requirements.txt

Inference

Predict and save segmentation results.

Requirements:

• model file: RESULTS/lobe/ZUNet.pth
• dicom: {subj_path}/dicom/
• in file: ProjSubjListDCM.in (if you want to run multiple cases)

# Infer one case (lobes)
python run_inference --mask=lobes --model=ZUNet --subj_path=25 --parameter_path=RESULTS/lobes/ZUNet.pth --pp=True --multi_channel=True

# Infer one case (airway)
python run_inference --mask=airway --model=UNet --subj_path=25 --parameter_path=RESULTS/airway/UNet.pth --pp=False --multi_channel=False

# Infer a list of cases
python run_inference --in_file_path=TE_ProjSubjListDCM.in

Parameters:

• mask: [airway, lobes, lung, vessels]
• model: [UNet, ZUNet]
• subj_path: Subject folder which has a dicom folder
• parameter_path: path to the *.pth
• pp: postprocessing [True, False]

Outputs:

• {model}_{mask}.img.gz
• {model}_{mask}.hdr.gz

Test

Calculate segmentation accuracy using the mask from the 'run_inference.py' and ground truth.

Requirements:

• predicted mask: *.img.gz & *.hdr
• label: *.img.gz & *.hdr

Train

Data Preparation

0. Get VIDA_sheet.csv

• Download VIDA_sheet as a .csv from either google drive or one drive.
• save as VIDA_sheet.csv

1. Get ProjSubjList.in

cd utils
python get_ProjSubjList.py /d/ENV18PM/VIDA_sheet_20220120.csv /d/ENV18PM/ENV18PM_ProjSubjList_20220120.in /d/ENV18PM/ImageData

2. Convert dicom to analyze

python convert_dcm2img.py /d/ENV18PM/ENV18PM_ProjSubjList_20220120.in

Start training

python main.py

• mask: airway, vessels, lung, lobe
• model: UNet, ZUNet
• debug: True, False
• save: True, False
• lr: learning rate
• train_bs: train batch size
• valid_bs: valid batch size
• epochs: train epoch
• n_case: number of cases to use