Fig. 1.

Components of a deep neural network. An input medical scan is processed by a sequence of network layers in the so-called ‘backbone’ applying non-linear transformations whose parameters (or weights) are learned during training. This results in a complex feature representation at the penultimate layer which is then processed by the final prediction layer. The prediction layer assigns weights to each feature in the penultimate layer, aggregates the weighted features, and generates an output prediction. The prediction layer has the role of making the decision on what information is being used for making predictions.