Abstract
In routine breast cancer diagnostics, pathologists often review each case twice-first to determine the need for immunohistochemical (IHC) stains, and a second time to issue the final diagnosis-creating significant workload and delays. We present an artificial intelligence-based system designed to streamline this process by distinguishing invasive carcinoma from fibroadenoma and in situ lesions in Whole Slide Images of H&E-stained breast biopsies, enabling automatic IHC stain requests for these lesions, while abstaining on non-target or low-confidence cases. The system leverages a weakly supervised method, trained directly on final diagnostic labels without the need for manual annotations. It achieves over 91% sensitivity and specificity across histological types in internal validation and shows strong generalizability in two external pilot studies. In a real-world setting, the system could determine if IHC stains should be ordered (sensitivity and specificity > 96%) and which stains (sensitivity and specificity > 81%). Compared to both expert pathologists and state-of-the-art models, our model performs competitively. Designed for real-world deployment, it is fully integrated into the Digital Imaging and Communications in Medicine (DICOM) standard. In validation data, it could have saved up to 43 h of pathologists' time. Our model represents a scalable solution for more efficient diagnostic BC workflows.