Machine-learning-based image analysis algorithms improve interpathologist concordance when scoring PD-L1 expression in non-small-cell lung cancer

Abstract

Programmed death ligand 1 (PD-L1) expression on tumour cells is the only predictive biomarker of response to immuno-modulatory therapy for patients with non-small-cell lung cancer (NSCLC). Accuracy of this biomarker is hampered by its challenging interpretation. Here we explore if the use of machine-learning derived image analysis tools can improve interpathologist concordance of assessing PD-L1 expression in NSCLC.Five pathologists who routinely score PD-L1 at a major regional referral hospital for thoracic surgery participated. 13 NSCLC small diagnostic biopsies were stained for PD-L1 (SP263 clone) and digitally scanned. Each pathologist independently scored each case with and without the Roche uPath PD-L1 (SP263) image analysis NSCLC algorithm with a wash-out interim period of 6 weeks.A consistent improvement in interpathologist concordance was seen when using the image analysis tool compared with scoring without: (Fleiss’ kappa 0.886 vs 0.613 (p<0.0001) and intraclass coefficient correlation 0.954 vs 0.837 (p<0.001)). Five cases (38%) were classified into clinically relevant different categories (negative/weak/strong) by multiple pathologists when not using the image analysis algorithm, whereas only two cases (15%) were classified differently when using the image analysis algorithm.The use of the image analysis algorithm improved the concordance of assessing PD-L1 expression between pathologists. Critically, there was a marked improvement in the placement of cases into more consistent clinical groupings. This small study is evidence that the use of image analysis tools may improve consistency in assessing tumours for PD-L1 expression and may therefore result in more consistent prediction to targeted treatment options.