Publication

Abstract : Introduction This study evaluates whether daily MVCT-derived shifts from the Ruby modular QA phantom encode systematic geometric information that can predict patient setup errors across six degrees of freedom. Methods A retrospective dataset comprising 450 consecutive days of Ruby phantom MVCT scans was paired with clinically applied 6-DOF patient setup corrections from head and neck (50 patients), thorax (50 patients), and pelvis (35 patients). Phantom shifts were augmented with engineered temporal features describing drift, variance, and motion stability. Machine learning regression models were optimised using Bayesian hyperparameter search (Optuna) and trained using an 80/20 split. Performance was assessed using per-axis R2, mean absolute error, root mean square error, accuracy within ±0.5 mm/° and ±1.0 mm/°, and Bland–Altman analysis. Independent testing was conducted on an additional 50-day unseen phantom dataset. Results Translational setup errors were predictably related to phantom-derived features, with the strongest performance observed in head and neck and pelvic treatments. In head and neck cases, leading models achieved translational MAE of 0.35–0.45 mm, RMSE <0.7 mm, and R2 > 0.90, with accuracy exceeding 85% within ±0.5 mm/° and >95% within ±1.0 mm/°.Pelvic performance was slightly reduced but clinically acceptable, while thoracic predictions were weaker due to respiratory motion. Rotational corrections showed limited predictability across all regions. Bland–Altman analysis demonstrated minimal systematic bias and stable generalization on independent testing. Conclusion Daily Ruby phantom MVCT data contain actionable information on systematic translational setup error in helical tomotherapy. Machine learning enables repurposing routine QA imaging into a predictive tool for enhanced geometric insight without additional dose or workflow burden.