Clinical evaluation of super-resolution for brain MRI images based on generative adversarial networks

Abstract

In magnetic resonance imaging (MRI), reducing long scan times is an urgent issue that could be addressed with super-resolution (SR) techniques. Most of the SR networks using deep neural networks (DNNs) have been evaluated only based on numeric metrics, and the image restoration quality for individual lesions is not evaluated sufficiently. Here, we evaluated the performances of different types of SR networks using DNNs for brain MRI, in terms of diagnostic performance and image quality. We focused on comparing the performance between generative adversarial networks (GANs) and non-GAN networks. There was a trade-off in such restoration quality between GAN- and non-GAN-based SRs, with the GANs being more accurate in restoring images of anatomical structures but less accurate in restoring those of lesions; non-GANs showed the opposite tendency. The non-GAN SRs were preferable in terms of diagnostic performance and image quality. This result suggested that the evaluation of DNN performance for lesions might be changed drastically by adding a clinical evaluation perspective. The dependence of network architecture on network performance obtained in this study will provide guidance for future development of SR DNN for medical images.