Fig. 3

1) Data annotation protocol, comprising whole slide image (WSI)-level viable residual tumor percentage labels and patch-level binary classification of positive vs negative regions. 2) Two-phase training paradigm, first exploiting cross-entropy loss for labeled patches, followed by training with pseudo-labeled WSI patches determined by the pretrained model. 3) Exemplar clinical applications of the model for residual tumor percentage quantification and spatial mapping of potential viable regions. shows where model weights were frozen