Fig. 5
a Comparative assessment of residual tumor percentage by different analytical approaches versus expert consensus ground truth. b Assessment of model-predicted residual tumor percentages against ground truth measurements across two study cohorts. The performance remained notable on the external Cohort 2 test set, showing satisfactory generalizability of our approach to new histological specimens. c Scatter plots comparing viable tumor percentage predictions for example patients from human experts and the deep learning model, illustrating general concordance between automated and manual assessment methods