Virtual patient-specific QA with DVH-based metrics.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning in Healthcare

Background:

• Patient-specific quality assurance (PSQA) is crucial in radiotherapy to ensure accurate dose delivery.
• Current PSQA methods may not fully capture real-world treatment delivery variations.
• Predictive modeling offers a potential enhancement for pretreatment QA.

Purpose of the Study:

• To develop and validate a virtual pretreatment PSQA procedure using machine learning.
• To quantify the dosimetric effects of predicted delivery discrepancies for IMRT and VMAT.
• To assess the potential of a "delivery-free" analysis to improve PSQA.

Main Methods:

• Developed machine learning models (boosted tree for IMRT, bagged tree for VMAT) to predict delivery discrepancies from DICOM-RT parameters.
• Coupled prediction models with Monte Carlo dose calculation for dosimetric analysis.
• Trained and validated models on extensive IMRT and VMAT patient data.
• Clinically applied models to predict machine parameters and dosimetric effects, comparing them to trajectory log file data.

Main Results:

• High accuracy in predicting delivery errors, with R-squared values up to 0.989 for inter-field comparisons.
• Strong correlation between predicted and actual dosimetric changes (R-squared up to 0.966 for IMRT, 0.907 for VMAT).
• Validated models demonstrated reliable prediction of machine parameters and their dosimetric consequences.

Conclusions:

• The virtual pretreatment PSQA procedure accurately predicts radiation delivery discrepancies and their dosimetric impact.
• Machine learning models can anticipate dosimetric effects from treatment delivery files.
• This "delivery-free" approach shows significant potential for enhancing pretreatment patient-specific QA in IMRT and VMAT.