KEY POINTS
- This retrospective study included 1,360 patients with stage I–IVa nasopharyngeal carcinoma treated with intensity-modulated radiotherapy, with or without chemotherapy, between 2013 and 2017.
- Patients were divided into a training cohort of 951 and an internal test cohort of 409. Radiomic and three-dimensional deep-learning models were compared with models based on 12 clinical variables.
- Adding primary-tumour radiomic coarseness improved local-control discrimination from a concordance index of 0.51 to 0.60 (p = 0.02).
- A DenseNet-121 model combining clinical variables with merged primary-tumour and nodal masks achieved the best distant-control performance: concordance index 0.68 versus 0.66 for the clinical reference model (p = 0.01).
- CT-derived features did not significantly improve prediction of overall survival or progression-free survival. Only 21 local failures and 42 distant failures occurred in the test cohort.
CLINICAL TAKEAWAY
Routine planning CT may contain useful prognostic information beyond conventional staging, particularly for local and distant failure. However, the absolute gains were modest, local-control discrimination remained limited, and independent external validation is required before these models can guide treatment intensification or surveillance.