A feasibility study of deep learning prediction model for VMAT patient-specific QA

Area of Science:

• Medical Physics
• Radiotherapy Quality Assurance
• Artificial Intelligence in Healthcare

Background:

• Patient-specific quality assurance (QA) is crucial for verifying the accuracy of radiation dose delivery in VMAT.
• Current QA methods can be time-consuming and resource-intensive.
• Accurate prediction of QA metrics can streamline the radiotherapy workflow.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for predicting gamma passing rates (GPR) in VMAT patient-specific QA.
• To leverage dose data from treatment planning systems (TPS) and Monte Carlo (MC) simulations for enhanced prediction accuracy.
• To improve the efficiency and reliability of VMAT QA procedures.

Main Methods:

• Utilized data from 710 clinical VMAT plans measured with an ArcCHECK phantom.
• Recalculated doses using Pinnacle TPS and MC algorithms on ArcCHECK phantom images.
• Employed a four-layer convolutional neural network (CNN) for GPR prediction, evaluating performance with error metrics and ROC curves.

Main Results:

• The DL model achieved low Mean Absolute Errors (MAE) for GPR prediction across various gamma criteria (e.g., 1.1% for 3%/3mm).
• High correlation coefficients (up to 0.72) were observed between predicted and measured GPR.
• Excellent classification performance was indicated by Area Under the Curve (AUC) values reaching 0.93 for different gamma criteria.

Conclusions:

• The developed DL-based approach shows significant potential for improving VMAT patient-specific QA efficiency and accuracy.
• This method can serve as a valuable tool for reducing the workload associated with patient-specific QA.
• The study highlights the utility of integrating TPS and MC dose data with DL for advanced radiotherapy QA.