Motivation: Collecting in vivo training datasets for deep-learning-based MR Fingerprinting (MRF) reconstruction and tissue mapping methods can be challenging. Goal(s): We aim to shorten the MRF acquisition using deep learning methods without relying on training datasets. Approach: We propose a new self-supervised MRF deep learning framework that requires only undersampled k-space measurements. MRF dictionary is incorporated as a physics constraint to regularize the reconstruction in an efficient manner. Results: Our method shows promising results on 4x accelerated in-vivo MRF scans and phantom data, achieving ~6 mins post-processing time per scan without requiring ground truth tissue property maps or network pre-training. Impact: Our method achieves better results than conventional dictionary matching and is faster than current self-supervised MRF methods.
Liu et al. (Tue,) studied this question.