Tri2D-Net is the first deep learning network trained for directly estimating overall cardiovascular disease (CVD) risks on low dose computed tomography (LDCT). The corresponding paper has been published on Nature Communications.
- Python 3.7
- PyTorch 1.4
- Computing device with GPU
Noted that our code is tested based on PyTorch 1.4
This model was trained on the National Lung Screening Trial (NLST) dataset. The NLST is made publicly available by the National Cancer Institute. The detailed data information and the split of the NLST dataset used in the paper is contained in NLST_data_split.csv.
- Heart Detection: RetinaNet was used in our study for heart detection.
- Resize & Normalization: The detected heart region was resized into 128x128x128. The image was normalized with a range of -300HU~500HU.
BEFORE RUNNING THE CODE, PLEASE DOWNLOAD THE NETWORK CHECKPOINT FIRST.
The trained model can be downloaded through this link. Please download the checkpoint to the ./checkpoint folder.
To predict CVD Risk from an image, run:
python pred.py--pathpath of the input image. #Default:./demos/Positive_CAC_1.npy--iteriteration of the checkpoint to load. #Default: 8000
The model takes a normalized 128x128x128 numpy.ndarray as an input, i.e., each item in the ndarray ranges 0~1.
A real number in [0, 1] indicates the estimated CVD risk.
We uploaded 4 demos in the ./demo folder, including one CVD negative case and three CVD positive case. One of the CVD positive subjects died because of CVD in the trial.
The name of the file indicates its label and the CAC grade evaluated by our radiologists.
Please cite these papers in your publications if the code helps your research:
@Article{chao2021deep,
author = {Chao, Hanqing and Shan, Hongming and Homayounieh, Fatemeh and Singh, Ramandeep and Khera, Ruhani Doda and Guo, Hengtao and Su, Timothy and Wang, Ge and Kalra, Mannudeep K. and Yan, Pingkun},
title = {Deep learning predicts cardiovascular disease risks from lung cancer screening low dose computed tomography},
journal = {Nature Communications},
year = {2021},
volume = {12},
number = {1},
pages = {2963},
url = {https://doi.org/10.1038/s41467-021-23235-4},
}
Link to paper:
The source code of Tri2D-Net is licensed under a MIT-style license, as found in the LICENSE file. This code is only freely available for non-commercial use, and may be redistributed under these conditions. For commercial queries, please contact Dr. Pingkun Yan.