A Radiomics-Based Machine Learning Perspective on the Parotid Gland as a Potential Surrogate Marker for HPV in Oropharyngeal Cancer

SIMPLE SUMMARY: The human papillomavirus (HPV) is an important prognostic marker in oropharyngeal squamous cell carcinoma (OPSCC) due to its involvement in carcinogenesis. The presence of the virus itself or markers of a past infection are usually determined by tissue specimen. This work applies predictive machine learning models to objectified imaging features extracted from routine imaging of patients newly diagnosed with OPSCC to classify cases of HPV-positive and negative tumors. In addition to results and classification performance being in line with existing literature on imaging-based HPV classification, imaging features of the parotid gland could be shown to yield relevant information. Parotid gland imaging offers potential as a screening tool and may aid clinical decision making in cases of cancer with unknown primary in the future. ABSTRACT: Background: In treatment of oropharyngeal squamous cell carcinoma (OPSCC), human papillomavirus status (HPV) plays a crucial role. The HPV-positive subtype tends to affect younger patients and is associated with a more favorable prognosis. HPV-associated lesions have been described in the parotid gland, which is included in routine imaging for OPSCC. This work aims to explore the ability of an ML system to classify HPV status based on imaging of the parotid gland, which is routinely depicted on staging imaging. Methods: Using a radiomics approach, we investigate the ability of five contemporary machine learning (ML) models to distinguish between HPV-positive and HPV-negative OPSCC based on non-contrast computed tomography (CT) data of tumor volume (TM), locoregional lymph node metastasis (LNM), and the parotid gland (Parotid). After exclusion of cases affected by streak artefacts, 53 patients (training set: 39; evaluation set: 14) were retrospectively evaluated. Classification performances were tested for significance against random optimistic results. Results: The best results are AUC 0.71 by XGBoost (XGB) for TM, AUC 0.82 by multi-layer perceptron (MLP) for LNM, AUC 0.76 by random forest (RF) for Parotid, and AUC 0.86 by XGB for a combination of all three regions of interest (ROIs). Conclusions: The results suggest involvement of the parotid gland in HPV infections of the oropharyngeal region. While the role of HPV in parotid lesions is under active discussion, the migration of the virus from the oral cavity to the parotid gland seems plausible. The imaging of the parotid gland offers the benefit of fewer streak artifacts due to teeth and dental implants and the potential to screen for HPV in cases of an absent or unlocatable tumor. Future investigation can be directed to validation of the results in independent datasets and to the potential of improvement of current classification models by addition of information based on the parotid gland.