Figure 2.

Deep-learning network structure. A 2D U-Net-based architecture consisted of four down-sampling steps and four up-sampling steps was implemented. Each encoder stage was followed by 2 × 2 max-pooling for down-sampling, and each decoder stage was followed by 2 × 2 up-sampling convolutional layers. Every stage incorporated two series of 3 × 3 2D convolutions, batch normalization, and rectified linear units (ReLU). Input images included T1- and T2-weighted images concatenated as two channels. The output image was an estimated T2 map. An L1 loss function was used.