Artificial Intelligence Based Strategies for Data-Driven Radial MRI

Magnetic Resonance Imaging (MRI) represents one of the major imaging modalities for its low invasiveness and versatility. Functional MRI (fMRI) need rapid acquisition to follow dynamic processes. Radial directions, due to their quality of overfitting the k-space center, are particularly suitable to average the effects of movements and are hence often used in fMRI. The acquisition of a complete dataset could be too slow to follow dynamic processes and undersampling is required to improve temporal resolution. Undersampling is accomplished randomly, basing on Compressed Sensing (CS) constraints to reduce aliasing and structured artifacts after reconstruction. However, reconstruction could be improved and/or data can be further reduced if a-priori information is collected regarding the underlying image under reconstruction and Artificial Intelligence (AI)-based strategies are used to drive the process. As an example of the effective synergy between AI and data-driven acquisition/reconstruction in radial MRI, we present a GReedy Adaptive Data-driven Environment (GRADE) for intelligent radial sampling that uses the power spectrum of the reconstructed image and AI-based superresolution strategies in an iterative acquisition/reconstruction process. A detailed description of the method is furnished and experimental results are reported. Results demonstrate that GRADE reduces data redundancy and converges first to high quality images with respect to other undersampling radial modalities, such as regular sampling and golden angle (GA).