Fig. 3

Diagram of the training workflow and the 2D residual U-Net architecture. Data was augmented to improve generalization of the network. Input image slices pass through three downsampling layers followed by a bridge layer and four upsampling layers. Each block in the Gadgetron U-Net diagram represents two cycles of batch normalization, convolution layer, and leaky ReLU activation. The final layer was a 1 × 1 convolution used to convert the output channels into a probability map, and no nonlinearity function was added after the output convolution. The U-Net outputs a probability map, which was then thresholded to create 2D lung segmentation masks