Deep Feature-Driven SVM Modeling for Arc-Level Quality Assurance In Prostate Radiotherapy
Abstract
Purpose
The aim of this work is to predict the arc level gamma passing rate (GPR) values for IMRT, VMAT, and FSRT plans at the prostate site using machine learning-based models using the SVM algorithm.
Methods
Here Support Vector Machine (SVM), a supervised max-margin model is used for regression analysis between dose and GPR. A total of 175 prostate treatment plans, comprising 470 arcs, were analyzed. Dose planes originally sized at 1280×1280 for IMRT and FSRT and 270×270 for VMAT were resampled to 512×512 and normalized to ensure consistent spatial representation. A pretrained ResNet model with frozen weights was used to extract deep features. ResNet captures the dose gradient characteristics and spatial modulation patterns. The original high dimensional feature set was reduced to 36 informative predictors using a multistage feature selection pipeline that included variance thresholding, correlation filtering, mutual information scoring, and recursive feature elimination. A polynomial kernel SVM model with 10-fold cross validation was trained using these features. Root mean square error(RMSE), mean absolute error(MAE) , coefficient of determination, and Pearson correlation coefficient were used to assess the model's performance.
Results
. Strong agreement between predicted and measured gamma passing rates was demonstrated by the model's RMSE of 2.03 percent, MAE of 1.09 percent, R2 of 0.816, and correlation of 0.909. 82.6% and 88.7% of predictions, respectively, fell within the corresponding tolerance bands under the 2% and 3% gamma criteria.
Conclusion
These results show that the suggested model offers clinically significant accuracy and consistently captures arc level variations. The machine learning model provides a way to lessen the workload associated with patient-specific quality assurance, and further research will expand this approach to other anatomical sites to facilitate the creation of a universal prediction model.