Non-invasive CT-based Deep Learning for Human Papillomavirus Status Prediction in Oropharyngeal Cancer

Human papillomavirus (HPV) status is a critical biomarker for treatment planning and patient management in oropharyngeal cancer (OPC). The purpose of this study was to develop a non-invasive and rapid alternative that achieves performance comparable to established clinical standards while enabling HPV classification without requiring additional examinations beyond the routine diagnostic and treatment workflow.

We employed a siamese neural network framework with 3D DenseNet backbones in each branch, trained in an end-to-end manner. To further enhance model performance, tumor mask volumes and frequency-domain representations of CT images were incorporated as additional modalities. A total of 1612 valid cases from three independent datasets were used to evaluate model performance, including internal validation on a large-scale dataset and three external validation experiments to ensure robustness and generalizability.

The proposed methods achieved an average AUC of 0.875 (95% confidence intervals (CI), [0.863,0.887]), an average Precision of 0.81 (95% CI, [0.8-0.82]), an average Recall of 0.89 (95% CI, [0.867,0.913]), and an average AUPRC of 0.922 (95% CI, [0.91,0.934]), proposed method achieved best result in all four metrics compared with reference methods. Pre-trained model achieved average AUC of 0.767 (95% CI, [0.735,0.781]), 0.726 (95% CI, [0.709,0.741]) and 0.839 (95% CI, [0.821,0.856]) in three external validations.

The proposed method achieved SOTA performance among CT-based HPV prediction methods for OPC and demonstrated performance comparable to established clinical methods. Furthermore, in accordance with FDA guidelines for Software as a Medical Device, the proposed radiogenomics approach demonstrated "good-enough" performance to support patient-informed decision-making in future clinical trials.