MDRANet: A multiscale dense residual attention network for magnetic resonance and nuclear medicine image fusion. (February 2023)

• Record Type: Journal Article
• Authors: Fu, Jun; Li, Weisheng; Peng, Xiuxiu; Du, Jiao; Ouyang, Aijia; Wang, Qian; Chen, Xin
• Is Part Of: Biomedical signal processing and control. Volume 80:Part 2(2023)
• Publication Date: 2023-02
• Subjects: Magnetic resonance and nuclear medicine -- Multimodal medical images -- Multiscale dense network -- Multiscale residual attention network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104382
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: MDRANet is proposed and applied to magnetic resonance and nuclear medicine image fusion. Four loss functions are used to improve the image fusion quality, and the gradient difference loss is proposed to enhance the edges and details of fusion images. The ablation experiments of four loss functions are performed, and the influences of different loss functions on the fusion results and residual images are analyzed. Abstract: Magnetic resonance and nuclear medicine images are the two categories of multimodal medical images. Magnetic resonance images reveal physiological anatomical information of patients, and nuclear medicine images accurately show tissue lesion information. Through medical image fusion algorithms, these fusion images containing both tissue lesion information and physiological anatomical information are obtained to provide sufficient information for clinical medical technologies. However, most existing fusion algorithms are based on mathematical transform domains, and these fusion results have the weaknesses of blurred edges, color distortion and detail loss. To address these problems, a multiscale dense residual attention network (MDRANet) is proposed and applied to magnetic resonance and nuclear medicine image fusion. MDRANet combines multiscale dense network and multiscale residual attention network to extract and enhance deep features. Moreover, four different loss functions are used to optimize MDRANet and improve the fusion quality. The experimental results show that the fusion results of our proposed algorithm have richer details and better objective metrics compared with the reference algorithms.