Introduction: The management of Oropharyngeal Squamous Cell Carcinomas (OPSCC) significantly benefits from improved stratification methods, especially considering the pivotal role of Human Papillomavirus (HPV) association in prognosticating disease course. Traditional methods, including p16 immunohistochemistry and HPV DNA detection, may not fully capture tumor heterogeneities and patient subpopulations.
Material and Methods: In this retrospective, multi-institutional study of 906 patients, we developed OPSCCnet, a Deep Learning (DL) algorithm to analyze standard Hematoxylin and Eosin (H&E) stained slides. Our methodology utilizes a Feature Pyramid Network with a ResNet-18 encoder for semantic segmentation of tumor areas and HPV association classification.
Results: OPSCCnet demonstrated a mean area under the receiver operator curve (AUROC) of 0.83 for the test cohort, which improved to 0.88 with a variance threshold for HPV-positive class probability. In prognostication, OPSCCnet showed a superior five-year survival rate prediction (96% [95% CI = 90–100%]) compared to standard HPV testing (80% [95% CI = 71–90%]). The three-tier threshold analysis using OPSCCnet indicated a high hazard ratio (HR = 0.15 [95% CI = 0.05–0.44]) in multivariate analysis, outperforming HPV testing (HR = 0.29 [95% CI = 0.15–0.54]).
Conclusion/Discussion: OPSCCnet, by analyzing standard H&E histological slides, outperforms traditional p16/HPV-DNA testing in patient stratification and survival prediction in OPSCC. These findings underscore the potential of AI-driven methodologies in enhancing patient stratification and advancing personalized medicine in oncologic care.
In dieser Studie erfolgte keine direkte externe Finanzierung für Forschung oder Manuskripterstellung. Jedoch erhielten wir für die zugrunde liegende Datenbank Fördermittel von Merck Sharp & Dohme. Diese Unterstützung könnte als potenzieller Interessenkonflikt angesehen werden, hatte jedoch keinen Einfluss auf die Studienergebnisse, die Dateninterpretation oder die Ausarbeitung dieses Manuskripts.