4× Super‐resolution of unsupervised CT images based on GAN

Abstract

Improving the resolution of computed tomography (CT) medical images can help doctors more accurately identify lesions, which is important in clinical diagnosis. In the absence of natural paired datasets of high resolution and low resolution image pairs, we abandoned the conventional Bicubic method and innovatively used a dataset of images of a single resolution to create near-natural high-low-resolution image pairs by designing a deep learning network and utilizing noise injection. In addition, we propose a super-resolution generative adversarial network called KerSRGAN which includes a super-resolution generator, super-resolution discriminator, and super-resolution feature extractor to achieve a 4× super-resolution of CT images. The results of an experimental evaluation show that KerSRGAN achieved superior performance compared to the state-of-the-art methods in terms of a quantitative comparison of non-reference image quality evaluation indicators on the generated 4× super-resolution CT images. Moreover, in terms of an intuitive visual comparison, the images generated by the KerSRGAN method had more precise details and better perceptual quality.