Abstract
BACKGROUND: Accelerated MRI reconstruction techniques are necessary to avoid long cardiac exams. K-space-based parallel imaging (PI) reconstruction has recently been adapted to deep learning with a scan-specific training technique entitled scan-specific robust artificial neural-networks for k-space interpolation (RAKI), which incorporates nonlinearity by applying convolutional neural networks. While the scan-specific aspect alleviates the need for a large training database, as it consists of a single-shot training, it consequently increases the overall reconstruction time. PURPOSE: The aim of this study is to adapt RAKI reconstruction to cardiac cine acquisitions by optimizing the training strategy and exploiting the spatio-temporal redundancy while ensuring image quality. METHODS: Ten fully sampled multi-slice cine data from the public cardiac OCMR database were used to compare the proposed method to standard reconstruction techniques (GRAPPA, RAKI, and rRAKI). To accelerate the reconstruction, the RAKI algorithm was simplified by removing the nonlinear activation units and reducing the number of layers, making it a parallelized GRAPPA-like reconstruction with only one convolution layer. Training of the weights was further accelerated by training only specific slices and cardiac phases of the whole cine stack of images. Image quality metrics such as the PSNR, NMSE, and SSIM were computed to evaluate the image quality, while the reconstruction time was also assessed. RESULTS: Quality metrics showed comparable results to state-of-the-art methods, while the average reconstruction time was reduced by 40 on average compared to GRAPPA, RAKI, and rRAKI. CONCLUSIONS: The reconstructions for the proposed method showed comparable image quality to standard methods while being significantly faster. Some "striping" artifacts remain with our method, which seem to be directly linked to the k-space-based optimization process.