Abstract / Summary
Background: Longitudinal imaging during carbon-ion radiotherapy (CIRT) may capture dose-related tumor evolution before clinical assessment. We investigated whether serial CT-based Δ-radiomic features follow treatment-fraction (dose escalation)-associated trajectories with prognostic relevance in patients with head and neck cancer (HNC) treated with CIRT. Methods: We analyzed 167 HNC patients receiving 16-fraction CIRT. Features were extracted from plan-CT, confirm-CT, and weekly-CT scans after 5, 9, and 13 fractions. Δ-radiomic features were defined as time-normalized relative changes from confirm-CT. Linear mixed-effects models, paired Wilcoxon signed-rank tests, and monotonicity assessments identified features with systematic longitudinal changes. Endpoints were progression-free survival (PFS), overall survival (OS), and local recurrence (LR). Optimal prediction timing was determined using optimism-corrected C-indices. Bootstrap resampling and multivariable Cox models assessed cutoff stability and incremental value beyond clinical factors. Results: During a median follow-up of 15.2 months (range: 1.2–67.7), 64 PFS events, 18 deaths, and 24 local recurrences occurred. Five, eight, and six candidate Δ-radiomic features were retained for PFS, OS, and LR. Selected prediction time points spanned fractions 5, 9, and 13. In multivariable analyses, four PFS-related features met the predefined robustness criteria and improved discrimination over the clinical model (C-index, 0.646–0.661 vs. 0.580; likelihood ratio test, p = 0.002–0.004). Sequential adjustment retained two OS-related and two LR-related features. Exploratory histology-specific analyses provided some support, primarily for two PFS-related features. Conclusions: Serial CT-based delta-radiomics identified treatment-fraction-associated imaging features associated with clinical endpoint risk after CIRT for HNC, particularly PFS. These findings are exploratory and require prospective external validation before clinical application.