Feasibility of Neural Network-Assisted T2 Mapping Utilizing Bloch Equation Simulations for Training Dataset Generation

Quantitative magnetic resonance imaging (MRI) aims to provide sensitive biomarkers that evaluate disease severity and progression rates. Spin-spin relaxation time of water protons (T2) is one of the most clinically valuable metric in quantitative MRI. Various techniques can be utilized for T2 mapping. One of the most accurate technique is based on the generation of a dictionary of simulated echo modulation curves (EMCs), using Bloch simulations, and searching the best match of experimental EMCs with the dictionary ones. However, the matching time still can be dependent on the size of the dictionary. The substitution of matching procedure by using of a pretrained neural network may reduce the dependency of T2 mapping time on the size of a dictionary. In this study, we aim to demonstrate the feasibility of a T2 mapping method based on neural network trained on a dictionary of simulated echo modulation curves obtained using Bloch simulations. The neural network was trained on simulated data augmented with noise. The study showed that augmentation with Rician noise provided lower absolute median error of T2 estimation (0.9%) than in case of using Gaussian noise (2.2%) or without noise (1.8%). These findings suggest that training a neural network using an EMC dictionary and using it for prediction of T2 values based on experimental EMCs can be feasible for quantitative MRI applications.