Fig. 1. Deep-learning framework for training and evaluating the model to classify and predict mutation.

Patients from TCGA were randomly divided into training cohorts (training and test) and internal validation cohort. Some patients had multiple virtual slides, and each slide was sliced into smaller “tiles”. The training, test, and internal and external validation sets were made up of multiple tiles from related cohorts. Model selection was done based on the performance in the test set. After learning and selection, the model was applied to tiles in the internal and validation sets to assess their performances.