AsymmetryNet: A Clinically Inspired Asymmetry Attention Model for Predicting HPV Status in Oropharyngeal Squamous Cell Carcinoma on Computed Tomography

Abstract Accurate, non-invasive HPV status prediction in oropharyngeal squamous cell carcinoma (OPSCC) is essential for risk stratification and treatment planning, yet current imaging methods lack sufficient sensitivity. We developed a novel framework combining 3D Masked Autoencoder (MAE) self-supervised learning (SSL) pretraining on a Swin Transformer encoder with 33 slice-aggregated asymmetry features (airway effacement, mass effect, midline shift, hypodensity metrics). In a multiinstitutional cohort (n = 173) using five-fold cross-validation and holdout testing, the combined model achieved CV AUC 0.909 (95% CI 0.850–0.957) and holdout AUC 0.828 (95% CI 0.643–0.964), outperforming ablations (SSL-only: 0.703/0.636; asymmetry-only: 0.545/0.588). At a highsensitivity threshold, it reached 100% sensitivity and 100% NPV. This establishes a new benchmark for imaging-based HPV prediction in OPSCC without segmentation, enabling reliable non-invasive risk assessment.