Fig. 1.

SöS-BC-4 and SCANB cohort descriptions and splitting criteria. a. The SöS cohort was first split into the training, internal test set 1, and internal test set 2 on the patient level. The split was stratified by clinical histological grading (NHG), ER status, HER2 status, and Ki-67 status. b. A five-fold cross validation (CV) split was further generated on the patient level within the training set (n = 1695 WSIs). Each CV fold consisted of a CV training set (80%) and a CV test set (20%) balanced on clinical NHG. The CV training set is further sub-split into the feature extractor training set (40%), the attention module (32%), and the tuning set (8%). c. SCANB cohort was used as the independent external test set