Insights on Scan-Specific Deep-Learning Strategies for Brain MRI Parallel Imaging Reconstruction.

Scan-specific deep learning strategies have been proposed for parallel imaging reconstruction in which auto-calibrated signals (ACS) are used for training. Here, we introduce methods to objectively optimize architecture and training details. In addition, we introduce a new metric to better characterize the quality of the reconstructed images. Various brain MRI situations are considered. The evaluated models encompass single-layer and three-layer residual CNN architectures with real and complex convolutions. Hyperparameters such as the level of linearity in leaky activation functions, loss function, kernel sizes and depths are optimized using grid-search with K-fold cross validation. The performances regarding ACS reference size and mode are also evaluated. An innovative COrrelation-Based Residual Artifact Index (COBRAI) quantifying the level of structured residual artifacts is proposed. Qualitative and quantitative comparisons are performed both on the FastMRI and in-house multi-contrasts 2D data. The proposed objective grid-search strategy based on ACS successfully provided optimized hyperparameters, retrospectively validated by enhanced image quality metrics. Notably, it is shown that nonlinearities produce structured residual artifacts, and that, among the models tested, a three-layer residual linear CNN with complex implementation and a reduced number of parameters is more robust, particularly providing less structured artifacts with less training data, leading to larger acceleration rates. Deep-learning MRI parallel image reconstruction in the scan-specific approach can be optimized using grid-search with K-fold cross validation. It was successfully applied in various 2D brain MRI situations. The quantification of structured residual artifacts with COBRAI is a useful complementary characterization to state-of-the-art metrics, and it can be used to drive model selection.