Author: Chuan He, Anh H. Le, Iris Z. Wang 👨🔬
Affiliation: Roswell Park Comprehensive Cancer Center, Cedars-Sinai 🌍
Purpose: To develop a non-measured and DVH-based (NMDB) IMRT QA framework integrating machine learning (ML) to classify lung SBRT VMAT plans prone to delivery errors
Methods: 560 Eclipse AcurosXB lung SBRT VMAT plans were delivered on a TrueBeam over the past three years. Delivery error’s means and standard deviations (STDs) for MLC and gantry were categorized by velocity and gravity during log analysis, while machine calibration uncertainty was assessed through periodic QA data. Final uncertainties were calculated using error propagation, and Gaussian noise was introduced to the control point values in the perturbed plans. Differences in PTV F-scores, integrating coverage and conformity, were calculated between original and perturbated plans and used to benchmark the plan’s susceptibility to delivery errors. Features extracted from DICOM doses and plans included basic plan parameters, planned DVH metrics, dosiomics, and histogram-based features. Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) models were trained using an 80:20 train-test split, feature reduction through recursive feature elimination, and hyperparameter optimization using grid search. Model performance was evaluated using ROC AUC and average precision (AP) scores.
Results: Significant correlations included gantry error STDs with gantry offsets (CC=0.53), MLC error means with velocities (CC=0.99), and MLC error means with gravity vectors (CC=0.77). OAR DVH discrepancies between original and perturbated plans were minimal, while PTV metrics showed average changes of 3.2% (V100%) and 3.0% (conformity index). The PTV F-scores varied by 1.6% on average, with a threshold of 2.3% identifying vulnerable plans. ML models achieved ROC AUC scores of 0.97 and AP scores of 0.90 (SVM, ANN) and 0.91 (RF).
Conclusion: The NMDB IMRT QA framework effectively classifies vulnerable lung SBRT VMAT plans with ML integration. This approach eliminates the need for physical measurements, facilitating online adaptive therapy and providing early feedback during planning.