Deep Learning Predicts HPV Association in Oropharyngeal Squamous Cell Carcinomas and Identifies Patients with a Favorable Prognosis Using Regular H&E Stains.

PURPOSE: Human papillomavirus (HPV) in oropharyngeal squamous cell carcinoma (OPSCC) is tumorigenic and has been associated with a favorable prognosis compared with OPSCC caused by tobacco, alcohol, and other carcinogens. Meanwhile, machine learning has evolved as a powerful tool to predict molecular and cellular alterations of medical images of various sources.

EXPERIMENTAL

DESIGN

We generated a deep learning-based HPV prediction score (HPV-ps) on regular hematoxylin and eosin (H&E) stains and assessed its performance to predict HPV association using 273 patients from two different sites (OPSCC; Giessen, n = 163; Cologne, n = 110). Then, the prognostic relevance in a total of 594 patients (Giessen, Cologne, HNSCC TCGA) was evaluated. In addition, we investigated whether four board-certified pathologists could identify HPV association (n = 152) and compared the results to the classifier.

RESULTS

Although pathologists were able to diagnose HPV association from H&E-stained slides (AUC = 0.74, median of four observers), the interrater reliability was minimal (Light Kappa = 0.37; P = 0.129), as compared with AUC = 0.8 using the HPV-ps within two independent cohorts (n = 273). The HPV-ps identified individuals with a favorable prognosis in a total of 594 patients from three cohorts (Giessen, OPSCC, HR = 0.55, P < 0.0001; Cologne, OPSCC, HR = 0.44, P = 0.0027; TCGA, non-OPSCC head and neck, HR = 0.69, P = 0.0073). Interestingly, the HPV-ps further stratified patients when combined with p16 status (Giessen, HR = 0.06, P < 0.0001; Cologne, HR = 0.3, P = 0.046).

CONCLUSIONS

Detection of HPV association in OPSCC using deep learning with help of regular H&E stains may either be used as a single biomarker, or in combination with p16 status, to identify patients with OPSCC with a favorable prognosis, potentially outperforming combined HPV-DNA/p16 status as a biomarker for patient stratification.