FIGURE 2.

Schematic of U-Net structure. ResNet-101v2 was used as the backbone network. The number on the left of every box represents the size of image or feature map. The number above the box denotes the number of feature maps (channels). A dropout was applied before every convolutional block for generalization. Arrows represent the operators. The meanings of the abbreviations are as follows: Conv, convolutional layer; RB, residual block; Leaky, leaky ReLU ( $\alpha$ = .1); BN, batch normalization; MP, max-pooling layer; s, strides, and c, channel of convolutional layer.