Figure 2.

Workflow of deep convolution neural network (DCNN) analysis. (A) Network Structure Overview. Eight-frame sequence was the input of the Gated Recurrent Unit (GRU)-based module. Sixteen-frame spliced image, output of the GRU-based module, and the clinic variables were the inputs of the convolution neural network (CNN)-based module. (B) GRU-based module. The feature extracted by the CNN-based Extractor was fed into a two-stage cascade GRU to get a one-dimension output. (C) CNN-based module. In the training stage, a jigsaw puzzle generator was applied to randomly generate three different patch sizes of image inputs based on the 16-frame spliced image. Three generated image inputs and the original image were then fed into pipelines composed by Conv Blocks and fully connected (FC) Blocks, respectively.