Generalizability of Deep Learning System for the Pathologic Diagnosis of Various Cancers

The deep learning (DL)-based approaches in tumor pathology help to overcome the limitations of subjective visual examination from pathologists and improve diagnostic accuracy objectivity. However, it is unclear how a DL system trained discriminate normal/tumor tissues specific cancer could perform on other types. Herein, we cross-validated DL-based classifiers separately tissue slides cancers bladder, lung, colon rectum, stomach, bile duct, liver. Furthermore, compared differences between frozen or formalin-fixed paraffin-embedded (FFPE) tissues. Area under curve (AUC) for receiver operating characteristic (ROC) ranged 0.982 0.999 when were analyzed by same preparation modalities AUCs drop 0.476 0.439 different types applied. Overall, optimal performance be achieved only