Mechanistical Modeling of Radiation Response of Parotid In Head and Neck Cancer Radiotherapy
Abstract
Purpose
Radiotherapy (RT) for head and neck (HN) cancer frequently induces xerostomia due to radiation‑mediated parotid gland damage. Mechanistic modeling of parotid response may improve understanding of toxicity development and enable more biologically informed prediction. This study applies a voxel‑based mechanistic model to clinical data and evaluates its predictive performance against the Lyman–Kutcher–Burman (LKB) NTCP model across multiple post‑treatment timepoints.
Methods
A cohort of 585 HN cancer patients treated with definitive RT and presenting with grade‑0 xerostomia at baseline was analyzed. Parotid glands were segmented from planning CT images, and individualized three‑dimensional dose distributions and longitudinal xerostomia assessments (days 0–1000 post‑RT) were incorporated. The model explicitly represented parotid ductal architecture and assigned voxel‑wise stem, progenitor, and acinar cell populations. Ductal voxels were initialized with 5% stem and 95% progenitor cells, while remaining tissue comprised 20% progenitor and 80% acinar cells. Stem cells underwent asymmetric division, and progenitor cells could self‑renew or differentiate. To improve training stability, model calibration followed a two‑stage approach: short‑term parameters governing acinar dynamics were optimized first, then fixed while long‑term recovery parameters for progenitor and stem cell populations were estimated.
Results
The mechanistic model consistently outperformed the LKB model, achieving AUC values at least 0.02 higher across all follow‑up timepoints. Sensitivity analysis indicated that early toxicity was primarily driven by acinar cell radiosensitivity, whereas long‑term recovery (>6 months post‑treatment) was dominated by stem cell radiosensitivity. All estimated parameters were biologically plausible, and the ductal compartment emerged as a critical contributor to sustained glandular recovery.
Conclusion
This voxel‑based mechanistic NTCP model demonstrates improved predictive performance over the LKB model while enabling continuous, biologically interpretable simulation of parotid injury and recovery. By incorporating spatial dose distribution and cellular population dynamics, the framework provides mechanistic insight into radiation‑induced xerostomia and supports toxicity prediction beyond static summary metrics.