Prediction and classification of VMAT dosimetric accuracy using plan complexity and log-files analysis.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Volumetric Modulated Arc Therapy (VMAT) is a complex radiotherapy technique.
• Ensuring dosimetric accuracy in VMAT is crucial for effective cancer treatment.
• Patient-specific quality assurance (QA) is essential but can be time-consuming.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting VMAT plan dosimetric accuracy.
• To classify VMAT plans into 'pass', 'control', or 'fail' categories based on predicted accuracy.
• To assess the utility of modulation complexity score (MCS) and log file analysis for VMAT QA.

Main Methods:

• Analyzed 1302 VMAT arcs from 651 treatment plans using MCS and linac dynamic log-files.
• Employed kernel regression to predict individual gamma-analysis (γ) pass rates (γ%) and mean values (γmean).
• Developed multinomial logistic regression, Naïve-Bayes, and SVM models based on MCS for plan classification.

Main Results:

• ML models achieved high prediction accuracy for γ% residuals (2.1-2.2%).
• Classification performance (precision, recall, F1) exceeded 90% for all ML models.
• Identified optimal MCS thresholds for failed (0.130) and reliable (0.270) plans with high sensitivity and specificity.

Conclusions:

• Machine learning models accurately predict VMAT treatment dosimetric accuracy.
• ML provides an efficient tool to support patient-specific QA in VMAT.
• A complexity-based traffic light system can effectively flag VMAT plan quality.