CT based Intratumor and Peritumoral Features for Predicting Prognosis of Patients with Head and Neck Cancer after Chemoradiotherapy: Using a Features Fusion Model

Objective This study developed a fusion feature model integrating traditional radiomic (Rad) and deep-learning (DL) features from distinct peritumoral areas to predict outcomes in head and neck squamous cell carcinoma (HNSCC) patients after chemoradiotherapy. Methods Data from 576 HNSCC patients (TCGA: 77; Center 1: 242; Center 2: 257) were retrospectively analyzed. Three datasets were created: training (Dataset A), external validation (Dataset B), and prediction (Dataset C). Four VOIs (GTV, GTV + 1mm, GTV + 2mm, GTV + 3mm) were evaluated using Rad models. The optimal VOI was used to build DL models, and combined Rad-DL models were generated. Clinical parameters were incorporated to construct a survival prediction nomogram. Performance was assessed via AUC, Kaplan-Meier (KM), and Cox regression. Results GTV + 2mm emerged as the ideal VOI. The Rad-DL model achieved the highest AUC (0.77 [0.70–0.84]), outperforming standalone Rad (0.64 [0.55–0.72]) and DL (0.68 [0.61–0.75]) models. Integration with clinical data further improved the Rad-DL nomogram (AUC = 0.83 [0.77–0.89]). In the prediction cohort, patients with high Rad-DL signatures showed superior 3-year progression-free survival (PFS: 86.6% vs. 32.2%) and overall survival (OS: 92.0% vs. 48.3%) compared to low-signature groups. Conclusions The GTV + 2mm VOI optimally captures prognostic information. The Rad-DL model, combining radiomic and deep-learned features, enhances prognostic accuracy for HNSCC patients, suggesting potential clinical utility in treatment planning and personalized care.