Figure 1.

(a) MRI signal synthetic model in the brain is used to train the neural network; while in vivo MRI data is used to test the neural network. Specifically, the proposed method assumed the feasible parameter space could be divided into boxes that were binary labeled for white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). (b) A deep neural network was used to perform pixel-wise estimation of T1, T2, B0 (not shown) and proton density/M0 (weighted by coil sensitivity variation) as a regression output, and WM, GM and CSF probabilities as a classification output. Each input pixel contained coefficients of principal components from a principal component analysis (PCA) preprocessing. The neural network contains 3 residual blocks, and each block has 8 fully connected layers, i.e. an MLP with “by-pass” connections to avoid gradient vanishing during training. The neural network was trained with T1 and T2 values from the parameter feasibility region as well as varied proton density weightings, varying noise levels, and binary class labels (in Fig. a).