2035 - Pretreatment Patient-Specific Quality Assurance Dose Prediction Model Based on Delivery Discrepancies of Linear Accelerator

Purpose/Objective(s): The accuracy of dose delivery in volumetric-modulated arc therapy (VMAT) plans is critically influenced by multi-leaf collimator (MLC) position and gantry angle deviations of linear accelerator. This study analyzes the relationship between these delivery discrepancies and 3D patient-specific quality assurance (PSQA) doses to ensure precise radiotherapy and improve patient safety.

Materials/Methods: Trajectory log files of VMAT plans from 1095 patients treated on a linear accelerator and 152 patients treated on a different model accelerator were analyzed. MLC position and gantry angle deviations between log files and treatment plans were assessed. Thirteen control point-level plan complexity metrics, encompassing leaf velocity (v_leaf), acceleration (a_leaf), gantry rotation velocity (v_gantry), leaf sequence variability (lsv) and so on, were calculated. The correlation between the plan complexities and delivery discrepancies was modeled using polynomial and first-order Fourier functions. These discrepancies were projected onto the patient's 3D volume to generate texture structures representing deviation impacts. These texture structures and the patient's planning dose were input into two Unet-based architectures with their deep and shallow features integrated to predict the 3D QA doses. Model performance was evaluated using R² for fitting, and SSIM and MAE (%) for 3D dose prediction accuracy.

Results: Deviations between planning and actual MLC positions reached 4 mm on the Clinical IX accelerator and 2 mm on the Edge accelerator. MLC position deviations showed strong correlations with v_leaf and a_leaf³ (R²>0.97 for both machine) at control point-level, the fitting intercepts exhibiting a 2p periodicity related to gantry angle. Gantry angle deviations varied randomly among treatment fractions but followed a definable distribution. These findings were consistent across different disease sites. The 3D QA dose prediction model achieved SSIM = 0.9844 and MAE = 0.6034.

Conclusion: This study establishes a robust control point-level prediction method for MLC position and gantry angle deviations in VMAT delivery. By incorporating these deviations into a 3D dose prediction, the model achieved a high accuracy, with potential applications in plan optimization and resource optimization.