关键词: 超高场磁共振, 并行成像, 欠采样图像, 复卷积, 深度学习

Abstract: Parallel imaging techniques can help overcome issues such as radiofrequency energy deposition, image inhomogeneity, and reduce scan time in ultra-high field Magnetic Resonance Imaging (MRI), while also lowering motion artifacts and accelerating data acquisition speed. To enhance feature extraction from MRI complex-valued data and reduce aliasing artifacts caused by undersampling in parallel imaging, a Residual Complex convolution scan-specific Robust Artificial-neural-networks for K-space Interpolation (RCRAKI) was proposed. The raw undersampled MRI scan data was taken as input by the algorithm, and the advantages of both linear and nonlinear reconstruction methods combined with a residual structure were leveraged. In the residual connection part, convolution was used to create a linear reconstruction baseline, while multiple layers of complex convolution were utilized in the main path to compensate for baseline defects, ultimately resulting in Magnetic Resonance (MR)images being reconstructed with fewer artifacts. Experiments were conducted on data acquired from a 7T ultra-high field MRI device developed by the Hefei Comprehensive National Science Center Energy Research Institute. The comparison results with the residual scan-specific Robust Artificial-neural-networks for K-space Interpolation (rRAKI) under a sampling rate of 40 Auto-Calibration Signals (ACS) and an acceleration factor of 8 for mouse imaging across different anatomical planes show that: in the sagittal plane, the Normalized Root Mean Square Error (NRMSE) of the proposed algorithm decreases by 59.74%, the Structural SIMilarity (SSIM) increases by 0.45%, and the Peak Signal-to-Noise Ratio (PSNR) increases by 13.04%; in the transverse plane, the NRMSE decreases by 7.97%, SSIM improves slightly by 0.001%, and PSNR increases by 1.09%; in the coronal plane, the NRMSE decreases by 35.03%, PSNR increases by 5.60%, and SSIM increases by 0.98%. Experimental results demonstrate that RCRAKI, combining complex convolution and residual structures, performs well across different anatomical planes of MRI data. The method can reduce the impact of noise amplification at high acceleration factors and reconstruct MR images with clearer details.

Key words: ultra-high-field magnetic resonance, parallel imaging, under-sampled images, complex convolution, deep learning