This is the code repository of paper Faint Features Tell: Automatic Vertebrae Fracture Screening Assisted by Contrastive Learning (arxiv).
In this paper, we propose a supervised contrastive learning based model to estimate Genent’s Grade of vertebral fracture with CT scans. Our method has a specificity of 99% and a sensitivity of 85% in binary classification, and a macro-F1 of 77% in multi-class classification. It can be concluded that forming feature space by contrastive learning could enhance CNN‘s capability of capturing faint feature and improve its performance on vertebrae fracture screening.
Our work has been accepted by BIBM2022 as short paper.
To support the research community of medical image analysis, our dataset is publicly available at OneDrive. You can find the detailed introduction in the dataset folder.
Additionaly, if you are going to train our model with your own data, you can refer to the pre-process folder to arrange the dataset.
Our training and validaion code can be found in the train folder, and information about our trained weight can be found in the model folder.
If you are going to use our code or data in your own work, we would be grateful if you cite our paper.
@inproceedings{wei2022faint,
title={Faint Features Tell: Automatic Vertebrae Fracture Screening Assisted by Contrastive Learning},
author={Wei, Xin and Cong, Huaiwei and Zhang, Zheng and Peng, Junran and Chen, Guoping and Li, Jinpeng},
booktitle={2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
pages={848--853},
year={2022},
organization={IEEE Computer Society}
}
Our dataset is under CC BY-SA 4.0 License, and our code is under MIT Lisence.