Deep-ERx2: Deep Learning Reconstruction for fast high-resolution non-Cartesian Compressed-Sensing MR Spectroscopic Imaging at 3T and 7T

Motivation: High-Resolution magnetic resonance spectroscopic imaging (MRSI) is a powerful, non-invasive method for detailed imaging of brain metabolism. However, traditional methods for reconstructing accelerated high-resolution whole-brain MRSI are time-consuming, posing challenges for its routine clinical application. Goal(s): Reconstruction of high-resolution whole-brain MRSI acquired at 3T and 7T in a timely fashion. Approach: Deep-learning reconstruction using recurring interlaced convolutional layers with joint dual-space feature representation for non-Cartesian Compressed-Sensing MRSI acquired by 3D ECCENTRIC sampling. Results: Deep learning ECCENTRIC reconstruction (Deep-ER) speeds up 600 times the reconstruction of high-resolution ECCENTRIC (k,t) data. Deep-learning reconstruction provides improved SNR metabolic maps across acceleration factors. Impact: Deep-ER enables high-resolution (3.4 mm isotropic) metabolic imaging with clinically feasible acquisition (4-9 min) and reconstruction times (1 min) at 3T and 7T. These times are compatible with the clinical workflow.