Predicting gamma passing rates for portal dosimetry-based IMRT QA using machine learning.

Area of Science:

• Medical Physics
• Machine Learning in Healthcare
• Radiation Oncology

Background:

• Intensity-modulated radiation therapy (IMRT) requires rigorous quality assurance (QA) to ensure accurate dose delivery.
• Portal dosimetry is a key component of IMRT QA, measuring dose delivery accuracy.
• Predicting IMRT QA outcomes can optimize the QA process and identify potential issues earlier.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting portal dosimetry-based IMRT QA gamma passing rates.
• To assess the accuracy and identify key features influencing gamma passing rates in IMRT QA.

Main Methods:

• Utilized 1497 beams from 182 IMRT plans across multiple treatment sites and linear accelerators.
• Characterized each beam using 31 features, including plan complexity and machine parameters.
• Trained and evaluated three tree-based machine learning algorithms (AdaBoost, Random Forest, XGBoost) using ten-fold cross-validation and a separate test set.

Main Results:

• AdaBoost and Random Forest models achieved 98% prediction accuracy within 3% of measured gamma passing rates (2%/2mm criteria).
• XGBoost showed slightly lower accuracy, with 95% of predictions within 3% of measured rates.
• All models identified plan complexity and beam aperture/jaw size as critical predictive features.

Conclusions:

• Tree-based ensemble learning models can accurately predict IMRT QA gamma passing rates.
• This machine learning approach enables proactive identification of IMRT QA failures.
• The findings support the development of more efficient and proactive QA strategies in radiation therapy.