Impact of Human Papillomavirus Infection And Tobacco Exposure On Oropharyngeal Squamous Cell Carcinoma Biology From A Radiomics Perspective

Abstract

It is well-established that human papillomavirus-associated (HPV+) oropharyngeal cancers (OPC) have distinct biology and favorable clinical outcomes in contrast to HPV-negative (HPV-) OPC. Though HPV+ OPC is currently regarded as one disease entity, there is growing evidence that extensive tobacco exposure can alter tumor biology and clinical outcomes, including overall survival. We hypothesized that these changes in the underlying biology would generate alterations in tumor phenotypes that can be captured by radiomics analytics. Data for biopsy-proven OPC patients dispositioned to definitive (chemo)-radiotherapy at a single institution between 2005 and 2012 were evaluated (n=319). Eligibility criteria included: concordant HPV DNA and p16 status as well as reported baseline patient reported smoking status in pack-years (PY). A total of 197 radiomics features were extracted from the gross volumes of the primary tumors. Redundant (Spearman correlation > 0.8) features were eliminated, and the remaining 37 features were used for modeling. Ensembles of random forests were used to correct for the imbalanced class representation in the data (275 HPV+ and 34 HPV-). All the models were internally validated using bootstrapping. Two classifiers of HPV status were trained using only radiomics features and then radiomics features in addition to smoking pack-years. Three-hundred and nineteen OPC patients were included in the analysis. The radiomic feature only classifier of HPV status produced an AUC = 70.5 (95%CI: 61.2-79.7). Including smoking PYs improved the AUC to 73.7 (65.1-82.2). Pack-years were also predictive in the second model (mean-decrease in accuracy: 2.0). Excluding ≥10 PY smokers (n= 110) from the HPV+ cohort resulted in superior discrimination between HPV+ (the subset with smoking PY<10; n= 175) and HPV- tumors using the same radiomic classifier, i.e. the largest AUC, 76.9 (95%CI: 67.3-86.4). The shape feature sphericity was the most discriminating feature for all three models (mean-decrease in accuracy: 2.4, 2.4, and 3.4). Other discriminating radiomic features included: Grey Level Co-occurrence Matrix features (Information measure correlation 1 and maximum probability), and shape feature ‘Convexity’. The 8th Edition of the AJCC Staging Manual is applied to all HPV/p16+ OPC tumors regardless of tobacco exposure history, presuming similar biological behavior and homogeneous clinical outcomes. Our radiomic data strongly suggest that conflating HPV+ OPC patients with significant tobacco exposure (>10 pack years) with HPV+ OPC patients without tobacco exposure may in fact be incorrect. Additional studies are required to explore the mechanisms which underlie the differential radiomic signature generated by tobacco exposure in HPV+ OPC.