Fig. 1.

Architecture of the U-Net deep CNN of this study. The deep CNN takes as input a $512\times 512$ image matrix. Solid arrows indicate the flow of the input matrix through the network, and dashed lines indicate merging of information through concatenation of feature maps. Convolutional layers are labeled as “Conv” followed by a triplet of numbers; the first number represents the number of feature channels of the layer, and the second and third number represent the height and width of the convolution window, respectively. Individual neurons were “dropped” at a probability of 0.5 in the two dropout layers of the network. All convolutional layers used the ReLU activation function, except where noted. Upsampling was acquired through nearest-neighbor interpolation.