基于生成对抗网络的 CT 图像无监督超分辨率分析

Improving the resolution of computed tomography (CT) medical images can help doctors to identify the lesion more accurately, which has important clinical significance. This paper studies how to obtain a low resolution image in the same domain as the real image by using a dataset containing only low resolution images without high-low resolution image pair data, and then constructs a training dataset close to the natural image pair by degrading the network and injecting noise. And a super-resolution generative adversarial networks including a super-resolution generator, a super-resolution discriminator and a super-resolution feature extractor (DeSRGAN) is designed to achieve X4 times super-resolution analysis of CT images. Experimental tests show that the DeSRGAN method is superior to the latest EDSR, RCAN, ESRGAN and other methods in quantitative comparison of X4 times CT images generated by super-resolution analysis without reference image quality evaluation indicators such as NIQE, BRISQUE and PIQE. At the same time, in terms of intuitive visual effects, the images generated by DeSRGAN method have clearer details and better perceptual effects.