Prediction of HPV-Associated Genetic Diversity for Squamous Cell Carcinoma of Head and Neck Cancer Based on $$^{18}$$F-FDG PET/CT

The relationship of human papillomavirus (HPV) to squamous cell carcinoma of the head and neck has been long known and explored. However, it has not been investigated whether cancer areas have the prognostic value of genetic diversity within HPV. Most existing studies in head and neck cancer analysis only involve a single imaging modality, e.g., computed tomography (CT) or magnetic resonance imaging (MRI), which may not provide complementary and diverse information for prediction task. Recently, positron emission tomography with 2-deoxy-2-[fluorine-18]fluoro-D-glucose integrated with CT ( $$^{18}$$ F-FDG PET/CT) has become a powerful imaging tool. In this study, we integrate these two imaging modalities (i.e., PET and CT) for HPV-associated genetic diversity prediction. Specifically, we design a deep 3D convolutional neural network (called PCNet) to learn PET and CT features in a data-driven manner, consisting of two branches (with each one corresponding to a specific data modality). The generated intermediate feature maps are further fed into a fully-connected layer for abstraction. Moreover, radiomic characteristics, which have been verified as a prognostic indicator in head and neck cancer, are concatenated with these data-driven deep learning features for final prediction. Experiments on 50 subjects demonstrate the effectiveness of our proposed PCNet. This is among the first attempts to explore the potential of PET/CT in differentiating genetic diversity in patients with HPV-associated squamous cell carcinoma of the head and neck.