Deep Learning Image Reconstruction of Accelerated 3D Cartesian Phase-Cycled bSSFP for Quantitative Mapping

Motivation: Phase-cycled bSSFP is valuable for quantitative mapping but requires acceleration methods with long reconstruction times. Goal(s): To explore deep learning networks for image reconstruction of accelerated phase-cycled bSSFP data. Approach: A U-Net and an end-to-end variational network were implemented and optimized on retrospectively and prospectively undersampled 3D Cartesian phase-cycled bSSFP data. Elliptical signal profiles and fat fraction maps were evaluated and compared against the ground truth and a compressed sensing reconstruction. Results: The end-to-end variational network performed similarly well as the compressed sensing algorithm on up to 8-times accelerated scans within 20s reconstruction time. Elliptical signal properties and fat fraction maps were well-preserved. Impact: The need for accelerated acquisitions in phase-cycled bSSFP significantly prolongs reconstruction times, necessitating efficient reconstruction techniques. This study demonstrates that deep learning achieves a feasible 20s reconstruction time while maintaining accurate fat fraction measurements in the knee of three volunteers.