Fig. 10.

Tile magnification optimization. Image tiles were extracted from slides in the training dataset at seven magnification sizes, ranging from 2.5x (tile width 1208 microns) to 40x (tile width 76 microns). Using the first training cross-fold, ten replicate models were trained at each magnification size using random weight initialization. a Area Under Receiver Operator Curve (AUROC) for models trained at varying magnifications. b Average Precision (AP) for models trained at varying magnifications