Fig. 7: Illustration of the deep-learning workflow for data processing and model evaluation.

We processed the WSI data by extracting tiles (1a), identifying tumor titles (1b), and generating small nonoverlapping tiles with color normalization. We selected key mutational genes (2) and identified biological pathways from mRNA (or CNA) expressions (3). Model training was based on a pretrained ResNet-101 model with an attention mechanism. After model selection, the trained model was used to test tiles and assess their prediction performances.