AANet: Artery-Aware Network for Pulmonary Embolism Detection in CTPA Images, MICCAI2022

pytoch==1.1.0 or 1.3.1 Newer version may need code adaptation.

simpleitk==1.2.4

tensorboardX

scikit-image

scikit-learn

tqdm

pandas

1. The vessel masks and lung masks are already open sourced in: AANet/PEData/CAD_PE_data/vessel The vessel masks are named like 001.nii.gz. The lung masks are named like 001_lungmask.nii.gz. Feel Free to use our vessel and lung annotation for your work. Just remember to quote us!

2. Download CAD-PE dataset from https://ieee-dataport.org/open-access/cad-pe Put CTPA images, e.g. 001.nrrd, in ‘AANet/PEData/CAD_PE_data/image’. Put PE labels, e.g. 001RefStd.nrrd, in ‘AANet/PEData/CAD_PE_data/label’.

3. Run nifty_preprocess.py to preprocess data. The code will use lung masks to crop the lung region ROI in the CTPA images and labels, and save them back to nifty file in ‘./PEData/processed_itk’. The image size will be smaller for faster loading in training.

Run train_aanet.py for training. Use the argument –unique_name to give a name. At the end of training, the code calls the inference.py automatically, and save the result in ‘AANet/pred_itk/unique_name_sth50’. You may want to train several times with different seeds. The official evaluation protocol of CAD-PE is a little bit unreasonable. If two near ground-truth PEs are detected by one connected predicted PE, only one ground-truth is counted as TP, and vice-versa. Therefore, small randomness, that causes two near PEs connected to one or one PE breaked to two, will cause relatively big difference in FROC curve.

Run evaluation.py to evaluate and plot FROC curve. Change pred_root in line 279 to the directory of your result, i.e. ‘AANet/pred_itk/unique_name_sth50’.

The guidance and the code may still have some small errors. Feel free to contact me by email or issue.