Habitat Radiomics Predict HPV Status in Oropharyngeal Cancer

This study developed a habitat-based radiomics classifier from CT images to predict HPV status in oropharyngeal cancer (OPC).

We analyzed pretreatment CT scans from 192 OPC patients. Tumor habitats were generated using a two-level unsupervised clustering approach, and radiomics features were calculated from both intratumoral and habitat-defined subregions. HPV ground truth was based on p16 IHC, and the ROSE algorithm was used to address class imbalance (85% HPV-positive). LASSO regression was used for feature selection. We developed and compared three separate models for statistical performance to predict HPV status: habitat radiomics classifier, intratumoral radiomics classifier, and combined radiomics classifier (habitats and intratumoral). Classifier performance was assessed using area under the receiver operating characteristic curve (AUCROC), and SHAP (SHapley Additive Explanations) was utilized to interpret features contributing to classifier predictions. Kaplan-Meier analysis was conducted to compare survival outcomes of HPV ground truth versus the radiomics classifier.

The habitat radiomics classifier significantly outperformed the intratumoral radiomics classifier, achieving AUCs of 0.970 (95% CI, 0.942-0.997) in the training cohort and 0.937 (95% CI, 0.843-1.00) in the test cohort. The combined radiomics classifier did not significantly improve performance. SHAP revealed that radiomic features of compact/spherical shape and uniform texture were associated with HPV-positive tumors, reflecting lower morphologic and textural heterogeneity than HPV-negative tumors. For the survival analysis, the HPV habitat radiomics classifier is indistinguishable from HPV ground truth in predicting overall survival (p = 0.52 for HPV-negative ground truth vs. HPV-negative classifier; p = 0.75 for HPV-positive ground truth vs. HPV-positive classifier).

The habitat radiomics classifier provides an imaging-based algorithm to predict HPV status in OPC by capturing spatial patterns of tumor heterogeneity that reflect biological differences between HPV-positive and HPV-negative tumors. This approach demonstrated strong performance, with additional support from SHAP-based interpretability and Kaplan-Meier survival analysis suggesting its potential clinical relevance.