Pre-Measurement Epid Image Prediction for IMRT Quality Assurance
Abstract
Purpose
Electronic portal imaging devices (EPID) are widely used clinically for patient-specific quality assurance (QA). Anticipating potential failures can help prioritize measurements to assess the need for plan revision and avoid downstream workflow disruptions. We developed a pre-measurement prediction method to identify plans at risk of QA failure before measurements.
Methods
We implemented a Residual Attention U-Net (RAU) architecture combining residual blocks for deep network training and attention gates for field-specific feature learning. Retrospectively, this study included 244 patients planned with 6X or 6FFF IMRT/VMAT and treated on 24 Linacs with the Millennium 120 MLC, either single-fraction or with the jaw size less than 2.5 cm by 2.5 cm. Treated disease sites include single or multiple lesion brain mets, oligometastases, lung, GI, and head and neck. 1118 fields from 223 patients were used for model training, and the remaining 21 patients’ 122 fields for testing. The calculated portal images direct from the QA plan were used as input, and the measured portal images served as ground truth. A rigid registration algorithm incorporating translation and rotation (±2°) aligned predictions to measurements for robust comparison. For the 21 test patients, mean absolute error (MAE) between the model-predicted image and the EPID measurement was quantified on jaw-masked field regions as a percentage of the corresponding QA plan field Dmax.
Results
Compared with the ground-truth EPID measurement images, the trained RAU predicted portal images had an MAE of 1.05 ± 1.80% for the independent test set. Prediction time was 0.43 ± 0.06 s per field on an RTX 6000 GPU.
Conclusion
Pre-measurement EPID image prediction using the trained RAU provides early feedback during treatment planning, enabling proactive workflow optimization. With accurate model-predicted portal images as the current focus, future work includes assessment of gamma analysis based on model-predicted images compared with that of measurement-based.