Fig. 12

Feature space visualization for the tile-based and multiple-instance learning models. a UMAP plot of post-convolutional layer activations, calculated using the final trained model, for all images in the University of Chicago training dataset. b Same as (a), but calculated for the TCGA test set. c UMAP plot of CTransPath features for all images in the University of Chicago training dataset. d Same as (c), calculated for the TCGA test set. e Mosaic map generated from the UMAP plot shown in (a). Three areas are magnified for closer inspection. Area 1 is enriched for HPV-positive images, Area 2 is in a zone of transition between HPV-positive and HPV-negative images, and Area 3 is enriched with HPV-negative images. Image tiles are shown using Macenko stain normalization