High-Resolution Image Estimation using Deep Learning
Xianjin Dai, Xiaofeng Yang · 2023
High-resolution medical imaging is highly desired in routine clinical applications either diagnosis or therapeutics. However, obtaining high-resolution images typically requires a long data acquisition time, which may be impossible under some circumstances due to the limited hardware capability, patient tolerance, and so forth. Thus, the trade-off among the spatial resolution, the speed of image acquisition, noise, and patient tolerance has to be made in each specific clinical situation. As advances in machine learning techniques, especially the take-off of deep learning in the past decades, generative adversarial networks (GANs), one of the most exciting innovations in deep learning, show a promising capability to generate synthetic yet realistic images. In this chapter, we present a self-supervised learning framework using cycle consistent generative adversarial network (cycleGAN), an improved GAN network. We validated the proposed approach for ultrasound (US), X-ray computed tomography (CT), and magnetic resonance imaging (MRI) high-resolution image estimation. Without any extra training data, neither paired nor non-paired images, high-resolution US/CT/MRI images were able to be estimated by using the intrinsic feature within the input images.