Deep-Learning Based Super-Resolution for Low-Dose CT

Recently we proposed a texture transformer network for the image super-resolution (TTSR), a reference-based deep learning super-resolution method, to improve the spatial resolution and suppress the noise of computed tomography (CT) images, where its effectiveness was demonstrated using the XCAT phantom. In this work, we focus on developing TTSR for real patient CT data. The mapping from the noisy low-resolution CT (LRCT) images to the clean high-resolution CT (HRCT) images is learned through TTSR. First, we iteratively reconstruct HRCT (full-dose, full projection data, 512x512 image resolution) and LRCT (quarter dose, 1/4 of all projection data, 128x128 image resolution) images using the CT projection data from 10 patients. Then, the TTSR models are trained using LRCT and HRCT images from nine patients to establish the translation from LRCT images to HRCT images. Finally, the SR model performance is quantitatively evaluated by peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) using the test patient images. Cubic spline interpolation and a generative adversarial network (GAN) with cycle consistency ("GAN-CIRCLE") are used for comparison. TTSR can restore more details in SRCT images than interpolation and GAN-CIRCLE. The quantitative results of PSNR and SSIM also show that TTSR can achieve the best performance. We finally show that TTSR using only 1/4 of all projection data can outperform LDCT with advanced denoising using all projection data. This development paves a way for deep learning-based SR for CT with lessened requirement on the number of detectors and much reduced computation time.