Fast GPU Implementation of a Scan-Specific Deep Learning Reconstruction for Accelerated Magnetic Resonance Imaging

Parallel Imaging is a technique commonly used in most clinical Magnetic Resonance Imaging (MRI) scans to mitigate the problem of long scan-times. In parallel imaging, information from multiple receiver antennas with different spatial sensitivities is combined to allow reconstruction from undersampled image information. Robust Artificial-neural-networks for k-space Interpolation (RAKI) has been recently proposed enabling parallel imaging reconstruction in MRI using convolutional neural networks (CNN) trained solely on a calibration signal corresponding to that image. While RAKI has demonstrated improved reconstruction performance compared to established techniques, its reconstruction time is prolonged due to the repeated application of the CNN, and the necessity of a training-phase for each receiver image. In this study, we propose an optimized RAKI implementation based on GPU parallel programming. The training phase duration is substantially shortened by optimizing the number of iterations and allowing for adaptively updated learning rates without compromising visual reconstruction quality. Efficient use of GPU resources was facilitated by a parallelized implementation of the training of multiple networks using CPU multiprocessing. The proposed implementation demonstrates more than 60-fold reduction in the reconstruction speed of clinical sample data compared with conventional sequential implementation, thus, easing the integration of RAKI in clinical applications.