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

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Healthcare

Background:

• Conventional patient-specific quality assurance (PSQA) relies on time-consuming physical measurements.
• Existing machine learning (ML) models predicting gamma passing rates (GPRs) have limited clinical utility due to weak correlation with dose-volume histogram (DVH) parameters.
• A novel non-measured and DVH-based (NMDB) PSQA framework is needed.

Purpose of the Study:

• To develop an NMDB PSQA framework using ML to identify treatment plans at risk of delivery errors.
• To improve the efficiency and accuracy of PSQA in radiation therapy.

Main Methods:

• Analysis of 560 lung SBRT VMAT plans using trajectory log files to categorize delivery discrepancies.
• Calculation and prediction of mean and standard deviation (STD) values for MLC and gantry positions.
• Integration of speed, gravity, and physical machine variability for error propagation.
• Recalculation of doses for perturbed plans and comparison of DVH metrics.
• Training and optimization of ML models (RF, SVM, ANN) using DICOM plan and dose features.
• Evaluation of model performance using ROC AUC and AP scores.

Main Results:

• High correlations observed between gantry/MLC errors and speed/gravity effects.
• While OAR DVH discrepancies were minimal (<1%), PTV metrics showed significant variations (e.g., V100% change of 3.2%).
• ML models achieved high classification performance (ROC AUC 0.97, AP 0.90-0.91) in identifying susceptible plans.

Conclusions:

• A novel NMDB PSQA framework for lung SBRT VMAT plans was successfully developed.
• The framework integrates DVH metrics (PTV F-score) and ML for real-time classification of susceptible plans.
• Eliminating physical measurements enables online adaptive therapy and early feedback, with significant clinical potential.