Figure 2.

Machine Learning Architecture and Flow Diagram.

The EfficientNetB4 architecture was used for the convolutional neural network with three dense layers with batch normalization and dropout. Ensemble models were generated by taking an average of the final predicted progression probabilities from the strongest individual PET and CT models. Features were extracted from the penultimate dense layer for the random survival forest model for overall survival risk. CT: computed tomography; PET: positron emission tomography.