Planning computed tomography radiomics predicted recurrence in human papillomavirus-positive oropharyngeal cancer

A multi-regional clinical-radiomics model reached an external-test area under the curve of 0.908 for 5-year recurrence prediction.

KEY POINTS

  • This retrospective multicenter study included 716 patients with human papillomavirus-positive oropharyngeal cancer from three public datasets: 390 in training, 166 in internal validation, and 160 in an independent test set.
  • The model used planning computed tomography radiomics extracted separately from the primary gross tumor volume and metastatic nodal gross tumor volume, then compared primary-only, physically fused, and feature-level fusion strategies.
  • The feature-level primary-plus-nodal strategy performed best; the XGBoost radiomics model reached area under the curve values of 0.907, 0.852, and 0.835 in training, internal validation, and independent testing.
  • The clinical-radiomics nomogram based on multi-regional features reached an area under the curve of 0.934 in training, 0.920 in internal validation, and 0.908 in the independent test set.
  • In the independent test set, the multi-regional combined model outperformed the primary-only combined model — 0.908 versus 0.857, p = 0.034.

CLINICAL TAKEAWAY

This study supports the idea that nodal radiomics may add prognostic information beyond primary-tumor radiomics in human papillomavirus-positive oropharyngeal cancer. For now, it is a risk-stratification research model, not a tool for treatment de-escalation or intensification decisions. The main limitations are retrospective public datasets, baseline imbalance in the external test set, manual contour dependence, absent radiotherapy dose data, and no prospective validation.

SOURCE

Radiation Oncology