Abstract

Info

Title:: Magnetic resonance image super-resolution reconstruction based on frequency-domain constraints and cross-fusion feature

Author(s):: LI Jiali; WANG Guozhong; ZHAO Haiwu; Artificial Intelligence Industry Institute, School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai 201620, China

Keywords:: Keywords: magnetic resonance imaging super-resolution reconstruction frequency domain constraint deep learning

Abstract:: Abstract: Magnetic resonance (MR) images are often used in clinical medical diagnosis, and high-resolution MR images are of high medical diagnostic value and can be recovered from low-resolution MR images using super-resolution reconstruction algorithms. The mainstream reference-based image super-resolution reconstruction algorithm can obtain fine image details, but still inevitably produces some artifacts. To address this problem, a frequency-domain constraints and cross-fusion feature (FCCF) model is proposed. The model introduces the frequency-domain loss function as a constraint and constructs a cross-fusion feature integration mechanism module to improve the quality of the generated images by cross-fusing image features of different resolutions, so that the reconstruction results have clearer details and no obvious artifacts. The experiment results show that the proposed method outperforms the existing super-resolution reconstruction algorithms when they are evaluated on synthetic and actual MR image data sets using peak signal-to-noise ratio and structural similarity.