De-noising of low-dose CT images using space-time nonlocal means over large-scale neighborhoods

Perfusion CT has been widely utilized as a functional imaging technology for the differential diagnosis of patients, and owing to its radiation-induced hazards, the radiation dose delivered to patients is always a significant concern. With lower health risks compared with high-dose CT, low-dose CT (LDCT) imaging, however, tends to be severely corrupted by quantum noise and streak artifacts. Accordingly, in this paper, we propose to improve the image quality by using the nonlocal means (NL-means) over large-scale neighborhoods, and by taking the information along the temporal axis into account, we further extend the proposed de-noising method from single 2D slice to volume. To lighten the computational burden, a CU-DA-dependent parallel computing is applied to accelerate the processing. Experiments on both phantom and clinical image datasets of different doses validate the excellent performance of the proposed method for improving the LDCT image quality.