A New Framework for 3D MR Fingerprinting with Efficient Subspace Reconstruction and Joint Posterior Distribution Estimation

Iterative image reconstruction of highly undersampled high-resolution 3D MR fingerprinting (MRF) is time-consuming and has high memory requirements. In this work, we propose to use stochastic gradient descent to accelerate the reconstruction and reduce the memory footprint. In addition, a conditional invertible neural network is used as a fast and flexible tool for estimating the posterior distribution of tissue properties from MRF. In a simulation study, we achieved an 11-fold and 45.5GB reduction in reconstruction time and memory requirement, respectively, compared with a conventional iterative method. Uncertainty maps of tissue properties derived from the estimated posterior distributions correlate well with reconstruction errors.