Self-supervised Dynamic CT Perfusion Image Denoising with Deep Neural Networks
A novel self-supervised denoising method was proposed for cerebral dynamic CT perfusion (CTP) images. CTP is widely used for stroke diagnosis but suffers from high noise and low resolution due to the dose constraints during dynamic scans. In this work, a denoising network was trained by mapping each time-frame image to the average of its adjacent neighbors. Because the noises in each time-frame image are independent, the network was trained to keep the underlying signals but remove the noises. The proposed method did not need any high-dose reference images for the training. Compared to conventional CTP denoising methods or networks trained using simulated dataset, the proposed method demonstrated more accurate quantification and higher spatial resolution.
Repository: https://github.com/wudufan/ctp_noise2noise
Related publications:
https://ieeexplore.ieee.org/abstract/document/9097940