Non‐measured and DVH‐based patient‐specific QA framework for lung SBRT VMAT with machine learning integration

Abstract Background Conventional patient‐specific quality assurance (PSQA) methods rely on time‐consuming physical measurements. While previous studies have successfully employed machine learning (ML) models to predict gamma passing rates (GPRs), their clinical utility remains limited due to GPR's weak correlation with dose‐volume histogram (DVH) parameters. Thus, developing a novel PSQA framework that is non‐measured and DVH‐based (NMDB) presents a promising alternative. Purpose To develop an NMDB PSQA framework incorporating ML to classify treatment plans susceptible to delivery errors. Methods This study analyzed 560 lung stereotactic body radiation therapy (SBRT) plans with volumetric‐modulated arc therapy (VMAT) delivered on a TrueBeam system (Varian Medical Systems), including 331 plans with saved trajectory log files. Log‐based analysis categorized delivery discrepancies in multileaf collimator (MLC) and gantry positions based on speed and gravity effect. Mean and standard deviation (STD) values for each category were calculated and predicted using linear regression. Additionally, physical variability was evaluated by analyzing periodic machine QA data to account for machine calibration uncertainty. Final discrepancies were calculated through error propagation, integrating speed, gravity, and physical variability. Gaussian noise was applied to control point values in all plans based on the estimated discrepancy means and STDs. Doses were recalculated for these perturbed plans, and the resulting DVH metrics were compared to the original plans. PTV F‐scores, combining coverage and conformality, were computed to quantify plan susceptibility to delivery errors, with a threshold set to classify the top 20% most vulnerable plans. ML models, including random forest (RF), support vector machine (SVM), and artificial neural network (ANN), were trained on features extracted from DICOM plans and doses, including basic plan parameters, planned DVH metrics, dosiomics, and histogram‐based features. The 80:20 train‐test split was implemented, with feature reduction based on statistical significance and collinearity. Models were optimized using hyperparameter tuning and recursive feature elimination. Their performance was assessed using the area under the receiver operating characteristic curve (ROC AUC) and average precision (AP) scores. Results High correlation coefficients (CCs) were noted between gantry error STDs and the gantry offsets (CC = ‐0.53), MLC error means and MLC speeds (CC = ‐0.99), and MLC error means and gravity vectors (CC = +0.77). Although DVH discrepancies for organs at risk (OARs) were minimal (< 1%), PTV metrics showed more considerable variations, including average changes of 3.2% for V100% and 3.0% for conformity index. PTV F‐scores varied by 1.6% on average, with a 2.3% threshold identifying susceptible plans. ML models demonstrated strong classification performance on the testing dataset, achieving ROC AUC scores of 0.97 and AP scores of 0.90 (SVM, ANN) and 0.91 (RF). Conclusion This study introduces a novel NMDB PSQA framework for lung SBRT VMAT plans, incorporating DVH metrics like the PTV F‐score and real‐time ML classification of susceptible plans. By eliminating the need for physical measurements, this framework enables online adaptive therapy and early feedback during planning, presenting substantial potential for clinical implementation and broader applications.