Predicting Adaptive Radiotherapy Session Duration Using Reference Plan Calculation Metrics
Abstract
Purpose
CT-guided online adaptive radiotherapy(CTgART) introduces substantial per-patient variability in session duration, complicating prospective scheduling. This study evaluates whether system-reported reference plan calculation metrics, specifically reference plan optimization and calculation times, can be combined into a predictive metric to prospectively assign adaptive patients to appropriately sized treatment time slots.
Methods
Reference calculation metrics (RCM), defined as the combined reference optimization and dose calculation times, were extracted from reference plan reports using a custom-built computer-use agent based on visual-language models (Fara-7b/Qwen3-8b). Session-level data was extracted from Varian Ethos treatment logs. A complexity model was developed using intent-specific median door-to-door (D2D) time with patient-level deviation adjustment. Fractions were classified as simple, moderate, or complex with prospective scheduling slots of 30, 45, and 60 minutes respectively. RCM was evaluated for correlation with adaptive D2D duration, session planning time, and session calculation time.
Results
A total of 187 CTgART patients treated between 2021 and 2025 were analyzed. Median RCM was 6.1[1.2–23.3]minutes, while median D2D time was 36.6[21.1–77.9] minutes. RCM demonstrated a positive correlation with adaptive D2D duration (r = 0.68). Session Calculation Metric time demonstrated a strong positive correlation with adaptive D2D duration(r = 0.95). Using the complexity model, 58 patients(32%) were classified as simple, 84 (44%) as moderate, and 45 (24%) as complex. RCM cutpoints corresponding to complexity bins were ≤2.6 minutes, 2.6–8.2min, and >8.2 minutes for simple, moderate and complex respectively. Compared to uniform 45-minute scheduling, the prospective model-based scheduling resulted in 8.2 minutes of unused time per fraction versus 11.4 minutes with uniform scheduling. Using the model-based approach,90.2% of sessions were appropriately allocated compared to 80.4% with uniform scheduling.
Conclusion
Reference calculation metrics provide a robust surrogate for adaptive treatment complexity and treatment time. Prospective scheduling informed by RCM improves time-slot utilization efficiency and allocation accuracy compared to uniform scheduling.