(610) Computed Tomography-based Classification of HPV Status in Head and Neck Squamous Cell Carcinoma

Introduction: Prognosis is better and TNM staging is different for human papilloma virus (HPV)-associated (HPV+) oropharynx cancer vs. non-HPV associated (HPV-) head and neck squamous cell carcinoma (HNSCC). Contrast enhanced CT (CECT) is part of the diagnostic workup for suspected HNSCC, therefore CECT-based HPV-status classification can provide this important information early, independently of pathology-based HPV-status ascertainment. We combined radiomics to extract quantitative data from CECT and applied machine learning to develop a method for determining tumor HPV status from CECT.

Methods: A classification model was developed in this retrospective analysis of 114 CECT from patients with confirmed HPV+ (n=27) or HPV- (n=87) HNSCC from the TCGA-HNSCC collection. An in-house developed artificial intelligence model segmented tumors and enlarged lymph nodes (eLN). Each image-mask pair was preprocessed via resampling to 0.5mm x 0.5mm x 2mm and clipped to [-50, 500] prior to radiomic feature extraction with PyRadiomics.
The cohort was separated into training (n=92) and testing (n=22) groups. 5-fold train-validation data splits were created, stratified by HPV, and eLN status . The scikit-learn Python package was used to develop a Random Forrest model using 214 total features from the primary lesion and eLN (Model A) to classify HPV status. Model sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated. An analogous classification model (Model B), built using radiomic features extracted from expert radiologist derived segmentations, was used for performance comparisons.

Results: Model A had a sensitivity for detecting HPV of 0.8, specificity of 0.94, PPV of 0.8, NPV of 0.94, and an accuracy of 0.91. Model B had a sensitivity of 0.4, specificity of 0.94, PPV of 0.67, NPV of 0.84, and an accuracy of 0.82.

Conclusions: Our model can accurately classify HPV status in patients with HNSCC from CECT. Further validation in larger datasets and prospective studies is necessary prior to implementing for clinical use.