Comparison of deep learning approaches to low dose CT using low intensity and sparse view data

Recently there has been considerable interest in using deep learning to improve the quality of low dose CT (LDCT) images. LDCT may be achieved by reducing the beam intensity, or by acquiring sparse-view data at full beam intensity. Additionally, if reducing beam intensity, one can consider denoising either the raw (sinogram) data, or the reconstructed image. We compare the performance of a convolutional neural network (CNN) in improving image quality using three approaches: denoising low-intensity images, denoising low-intensity sinograms prior to reconstruction, and denoising sparse-view images. Our results indicate that images produced from low-intensity data are superior to images produced from sparse-view data, after correction by the CNN. Additionally, in the low-intensity case, denoising in the sinogram or image domain provides comparable image quality.