Circumventing the Curse of Dimensionality in Magnetic Resonance Fingerprinting through a Deep Learning Approach

MR fingerprinting (MRF) is a rapid growing approach for fast quantitave MRI. A typical drawback of dictionary-based MRF is its explosion in size as a function of the number of reconstructed parameters, according to the curse of dimensionality. Deep Neural Networks (NNs) have been proposed as a feasible alternative, but these approaches are still in their infancy. We tested different NN pipelines on simulated data: we studied optimal training procedures by including different strategies of noise addition and parameter space sampling, to achieve better accuracy and robustness to noise. Four MRF sequences were considered, two of them designed to be more specific for B_1^+ parameter encoding: IR-FISP, IR-FISP-B_1, bSSFP and IR-bSSFP-B_1. A comparison between NN and the dictionary approaches was performed using a numerical brain phantom. Results demonstrated that training with random sampling and different levels of noise variance yielded the best performance. NN performance was greater or equal than dictionary-based approach in reconstructing MR parameter maps: the difference in performance increased with the number of estimated parameters, because the dictionary method suffers from the coarse resolution of the MR parameter space sampling. The NN approach resulted more efficient in terms of memory and computational burden, and thus has great potential in large-scale MRF problems.