Abstract 3294: Development and validation of an AI-based PD-L1 scoring algorithm in NSCLC samples

Abstract

Abstract Introduction: Lung cancer is the leading cause of cancer-related mortality worldwide, with non-small cell lung cancer (NSCLC) accounting for over 80% of all cases (1). The survival rate of patients with NSCLC is poor, with an overall 5-year survival rate of less than 30% (2). Programmed death-ligand 1 (PD-L1) expression on tumor cells is a response predictor to immune checkpoint inhibitor therapy and has become a cornerstone of first-line therapy for patients with advanced NSCLC without targetable alterations (3). However, poor inter- and intra-observer concordance when reporting PD-L1 expression might imply patients receive suboptimal treatment (4). Here, we present an AI-based algorithm for the clinical scoring of tumor PD-L1 expression in NSCLC samples to support pathologists and increase scoring consistency. Methods: HALO PD-L1 AI was developed and validated on a subset of routine diagnostic cases stained with the SP263 PD-L1 clone. To obtain the whole slide percentage of PD-L1 tumor positive cells and report the tumor proportion score (TPS), a ResNet 18-based network was trained with 141354 expert annotations to classify cells as ‘PD-L1 tumor positive’, ‘PD-L1 tumor negative’, or as ‘other’. Prior to PD-L1 cell classification, the algorithm automatically removes artifacts and segments the tissue into tumor and benign regions. The algorithm was validated by comparing the TPS score from three pathologists with the TPS score obtained from HALO PD-L1 AI on 203 whole slide images. Results: The three pathologists were in complete agreement in 64.0% of the cases. In pairwise comparisons, percent agreement ranged from 74.9% to 77.3%. Agreement of HALO PD-L1 AI with the pathologists’ mode was 73.4% overall, with specific agreement at the clinically relevant cut-offs ranging from 0.71 to 0.78. Intraclass correlation coefficient (ICC) between HALO PD-L1 AI and pathologists’ TPS scores was 0.95 (95% confidence interval 0.93 - 0.97). Conclusions: HALO PD-L1 AI is in agreement with pathologist TPS scores in routine diagnostic cases. Usage of HALO PD-L1 AI in a diagnostic setting has the potential to support pathologists scoring PD-L1, saving pathologists’ time and ensuring consistency in the reported results.