Speaker
Description
Given the continuous increase in modulated radiation therapy plans, patient-specific quality assurance became mondatory. Measurement-based patient-specific quality assurance is time-consuming and burdensome for medical physicists, particularly in busy radiotherapy centers. This study predicts and classifies gamma passing rate outcomes at the planning stage for head-and-neck volumetric modulated arc therapy treatment plans using texture features calculated from 2D dose distributions on cylindrical phantom. Four machine learning models namely, Random Forest, Decision Tree, XGBoost, and Support Vector Machine, were developed and trained on 2,428 planar dose distributions from 97 H&N cancer VMAT plans, using 100 texture features derived from the Gray Level Co-occurrence Matrix. γ3%/3mm, γ3%/2mm, and γ2%/2mm gamma passing rates, to would be predicted, were measured using the PTW OCTAVIUS-4D phantom. Results showed that the Random Forest model outperformed the others, demonstrating higher accuracy, precision, and Area Under the Curve, with lower errors and higher correlation cœfficient. Support Vector Machine had the highest prediction errors and lower rs compared to all others. Re-training the Random Forest model with the top 30 features further improved its performance in terms of prediction and classification. An AI-based desktop application was devloped and installed on TPS workstation to significantly facilitate the clinical workflow. Texture features, particularly contrast, were identified as key predictors of PSQA outcomes, highlighting the potential of combining contrast with random forest model for efficient PSQA in clinical practice.
| Abstract Category | Medical Physics |
|---|