Motivation: Standard deep learning approaches provide fast and accurate parameter estimation in magnetic resonance imaging (MRI) but still suffer from lack of network interpretation and sufficient training data. Goal(s): To propose one way that solely relies on the target scanned data and does not need a pre-defined training database with some Interpretability. Approach: We provide a proof-of-concept that embeds Bloch equation of MRI into the loss of physics-informed neural network (PINN). Results: PINN enables learning Bloch equation, estimating T2 parameter, and generating a series of physically synthetic data. T2 maps with phantom and realistic data obtained by PINN and least square are comparable. Impact: The proposed method provides a new way to quantify tissue parameter, which does not require analytical formula of Bloch equation under specific sequences, and is expected to simplify the sequence design of quantitative magnetic resonance imaging.
Cai et al. (Tue,) studied this question.