Abstract

Treatment decisions in oncology are commonly informed by the visual assessment of immunohistochemistry (IHC) biomarkers (such as PD-L1 expression) by pathologists. However, pathology services face mounting pressure as diagnostic demand increases and workforce decreases. Digital pathology and artificial intelligence have the potential to streamline the diagnostic workflow thereby improving pathologists’ workload, accelerating turn-around-times and facilitating access to testing. Here, we report the in-house development and analytical validation of a deep learning algorithm for automated scoring of PD-L1 expression in samples processed with the VENTANA PD-L1 (SP263) Assay. The algorithm was trained to score PD-L1 expression in tumour cells and in immune cells using 29318 manually annotated cells across a set of 150 PD-L1 IHC images from 30 urothelial carcinoma (UC) samples. The algorithm was then validated in an independent cohort of UC samples. In the validation cohort, the algorithm demonstrated high inter-scan reproducibility (99% overall percent agreement, N≤197), high inter-scanner reproducibility (100% overall percent agreement, N≤33) and substantial agreement with pathologist-based scoring of PD-L1 expression (84% overall percent agreement, N≤195). In conclusion, this study shows that our deep learning algorithm has favourable analytical characteristics to assist pathologists in scoring PD-L1 in both tumour cells and immune cells.