Multi-Scale and Multi-Contrast Magnetic Resonance Image Super-Resolution Reconstruction

Magnetic resonance imaging (MRI) can clearly show the structures of normal and pathological tissues, which helps doctors to make accurate diagnoses. However, due to the limited hardware, special imaging principles, and involuntary movements of patients, only low-resolution (LR) MRI images can be acquired within a limited time, which increase the probability of misdiagnosis. To address the issue, in view of the fact that MRI images with the same anatomical structures but different contrasts and resolutions are practically available, this paper proposes a multi-scale and multi-contrast MRI super-resolution (SR) reconstruction method, which utilizes the complementary information of auxiliary images to assist in the SR reconstruction of target LR images. Concretely, a dual-branch structure is adopted to extract multi-scale features from the target LR MRI image and the associated auxiliary image, and a Multi-scale and Multi-receptive-field Attention Fusion Block (MMAFB) is designed to fully utilize the texture and edge information of the auxiliary image. Furthermore, a Progressive Up-sampling based Multi-stage Feature Learning Module (PUMFLM) is used to produce a high-quality and high-resolution target image. Experimental results on the IXI and BraTS19 datasets in terms of PSNR surpass the second-best method by approximately 0.55 dB and 1.17 dB, respectively, which demonstrates that the proposed method outperforms existing single-contrast and multi-contrast SR reconstruction methods.