From Dose Plane to Gamma Passing Rate Using Convolutional Neural Networks, a Rapid QA for Online Adaptive Radiotherapy
Abstract
Purpose
To develop and evaluate a convolutional neural network (CNN) framework that estimates portal dosimetry gamma passing rate (GPR) directly from predicted dose plane images, enabling rapid quality assurance (QA) decision support for workflows relevant to online adaptive radiotherapy (OART), where conventional measurement-based QA is often impractical.
Methods
A retrospective dataset of 209 delivery fields from 87 prostate cancer patients treated with IMRT on a Halcyon linear accelerator was analyzed. Predicted portal dose planes generated by the treatment planning system were used as input to CNN models. Corresponding measured portal dosimetry QA (PDQA) gamma passing rates, evaluated using 1 mm/2% gamma criteria, served as ground truth. A stringent gamma criterion was selected to increase sensitivity to delivery and modeling variations, as standard clinical criteria (e.g., 2 mm/2% or 2 mm/3%) resulted in uniformly high passing rates with limited dynamic range. Model performance was assessed using three-fold cross-validation. Mean absolute error (MAE) was the primary metric. Root means square error (RMSE) and coefficient of determination (R²) reported as secondary metrics.
Results
Across cross-validation folds, the model demonstrated consistent agreement between predicted and measured gamma passing rates. The mean MAE was 2.76%, and the mean RMSE was 3.55%, indicating low prediction error across clinically relevant GPR ranges. The mean R² was 0.38, reflecting moderate correlation in the presence of delivery and measurement variability inherent to portal dosimetry. Prediction performance was stable across folds and included delivery fields with GPR values near institutional QA action thresholds.
Conclusion
This study demonstrates the feasibility of estimating portal dosimetry gamma passing rates directly from predicted dose plane information in clinical scenarios where measurement-based QA is not feasible, such as OART. The proposed approach provides rapid, quantitative QA decision support without additional measurements. With larger datasets and extension to additional treatment sites, broader clinical applicability is anticipated.