Author: Jenghwa Chang, Kuan Huang, Lyu Huang, Jason Lima, Jian Liu, Farzin Motamedi 👨🔬
Affiliation: Northwell, Department of Computer Science and Technology, Kean University, Physics and Astronomy, Hofstra University, Hofstra University Medical Physics Program 🌍
Title: A Deep Learning Approach to the Prediction of Gamma Passing Rates in VMAT Radiotherapy Plans for Adaptive Treatment.
Purpose: This study aims to develop a deep learning algorithm to predict the gamma passing rate (GPR) for measurement-based patient-specific QA in prostate/node volumetric modulated arc therapy (VMAT) plans. By leveraging prior clinical data, this tool can enable rapid QA predictions, offering a streamlined approach to support adaptive radiotherapy workflows.
Methods: A retrospective analysis of 100 prostate/node VMAT plans was performed. Verification plans were generated for the PTW Octavius 1500 phantom, and GPR values were calculated using the 2 mm/3% criteria with a 10% threshold. The dataset was split into training (80%) and testing (20%) subsets. Dose planes were extracted individually from each VMAT arc to preserve modulation information. The 2D dose plane was used as the model input, paired with corresponding GPR values as output. The ResNet34 architecture was customized for single-channel dose plane inputs, and Long Short-Term Memory (LSTM) layers were integrated to capture dependencies across sequential dose planes. The model was optimized with an MSE loss function. It`s performance was assessed using mean squared error (MSE), coefficient of determination (R²), and Pearson correlation coefficient, with results visualized via scatter plots predicted vs. measured GPR values.
Results: The final model achieved a mean squared error (MSE) of 1.0, an R² of 0.58, and a Pearson correlation coefficient of 0.9 between predicted and measured GPR values. These results demonstrate a strong positive correlation.
Conclusion: This deep learning model provides accurate predictions of GPR for prostate/node VMAT treatments, demonstrating its potential to replace traditional IMRT QA procedures. By offering rapid and reliable GPR predictions, this tool may enhance the safety and efficiency of adaptive treatment workflows. Future work will focus on external validation and extending this approach to other treatment sites.