Abstract 6172: Accurate PD-L1 IHC assessment in urothelial carcinoma by an AI algorithm robust across multiple sites, antibody clones, and scanners

Abstract

Abstract Background: Immunotherapy against PD-L1 is used for treatment of several indications of Urothelial Carcinoma (UC). Recent results indicate that it is likely to become part of the standard first-line treatment regimen for advanced UC in the near future. Accurate readout of Immunohistochemical (IHC) PD-L1 expression is therefore essential to inform UC treatment decisions. However, manual PD-L1-scoring is prone to high inter- and intra-observer variability. Artificial intelligence (AI) may help to improve standardization, accuracy and efficiency of PD-L1 IHC scoring. Yet, previous AI-based approaches struggled to show consistency across images that are diverse with regards to the source institutions, antibody clones, and scanning hardware used. Methods: We developed an AI software for PD-L1 tumor proportion scoring (TPS/TC) on whole-slide images (WSIs) of UC without human supervision. PD-L1 analysis by the software is fully-automated, does not require any adaptation steps, and is concordant with current clinical guidelines. The performance of the software was evaluated on two cohorts of PD-L1 stained UC samples against ground-truth (GT) TPS/TC scores given by human pathologists in clinical routine. The first cohort (n = 53) consisted of samples stained with the 22C3 antibody clone (Agilent) derived from four institutions and three scanner models. It contained both tissue microarray samples and scans of UC resectates. A second validation cohort consisted of n = 83 WSIs stained with the SP263 clone (Ventana), collected from two institutions. Results: For the threshold of TPS≥1% (cut-off established for adjuvant therapy with Nivolumab), agreement rates of AI scores with human GT scores were 92.5% for cohort 1 (22C3) and 96.4% for cohort 2 (SP263). Sensitivity/specificity values were 92.3%/92.9% for cohort 1 and 100%/93.6% for cohort 2. Across both cohorts (n = 136) the agreement rate between AI and human was 94.9%. Conclusions: Across 136 UC samples stained with two PD-L1 antibody clones, and derived from a total of four institutions and three scanner models, the fully automatic AI system showed excellent agreement with human pathologist scores. These results on challenging validation data show the potential of AI applications to standardize and optimize PD-L1 scoring and demonstrate consistency and safety of suitable AI-based systems for application in clinical routine. Citation Format: Niklas Abele, Diego Calvopiña, David Mulder, Patrick Frey, Emre Karakok, Fadime G. Salman, Murat Oktay, Markus Eckstein, Sebastian Springenberg, Arndt Hartmann, Tobias Lang. Accurate PD-L1 IHC assessment in urothelial carcinoma by an AI algorithm robust across multiple sites, antibody clones, and scanners [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 6172.