Super resolution of Covid-19 CT-Scan Images

Abstract

Acquisition of high quality CT images is difficult, because it requires exposing patients to high doses of radiation. Super resolution algorithms can help in over coming this problem and obtain higher spatial resolution in CT images. Much deep learning based architecture have been proposed in the literature to overcome this problem. We perform the task of super resolution on a U-Net and study the effects of 2 preprocessing methods which are scaling and zscore. The evaluation strategy for the super resolution of CT images in the literature uses the Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), however the results are published for the entire image. This is not a good practice for the evaluation of SR, we propose a novel region based similarity measurement practice and a lung specific or region of interest based similarity measurement. We further bifurcate the SSIM metric into it�s 3 component, i.e. luminance, contrast and structure, and study the impact of super resolution on each of these components.