Figure. 5
Neural network architecture. The deep convolutional neural network architecture for the AI system is made up of the feature encoder (green), the classification branch (yellow), and the localization branch (red). The feature encoder generates a latent representation of each radiograph. The classification and localization branches use the latent representations of radiographs to produce a probability model of expert radiologists identifying abnormalities for a category and a probability map of where the expert radiologists may identify abnormality for a category, respectively.
