Fig. 4

Network architecture of the DDSC-Net. Our DDSC-Net architecture includes multi-scale input of image pyramid and DDSC-blocks. As shown in the figure, each DDSC-block is composed of five “BN + Relu + depthwise separable convolution” dense connection layers, the convolution kernel is 3 x 3.The down sample is the maximum pooling layer, and the upsample is transposed convolution. Finally, the output of the network uses a softmax activation function to classify the output map into three categories: 0, 1 and 2. 0 is the background, category 1 is the optic disc, and 2 is the optic cup