An Adjustable Deep Learning Model for Improved EPID Dosimetry
X He*, F Li, Q Fan, W Cai, L Cervino, J Moran, S Lim, X Li, T Li, Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
MO-E115-IePD-F4-2 (Monday, 7/11/2022) 1:15 PM - 1:45 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 4
Purpose: The accuracy of analytical EPID dosimetry models deteriorate for highly modulated radiation beams. This study aims to investigate the feasibility of an accurate and robust portal dose image prediction (PDNet) model using deep learning (DL) and thousands of data measured for MLC shapes and jaw sizes, which can also be efficiently adapted to any treatment machine for a machine-specific model.
Methods: PDNet consists of a core module and a fully connected layer, where the core module is an attention U-net with a cost function that combines perceptual loss and multiscale structural similarity loss. In this study, 54 IMRT fields from six clinical plans were delivered on a TrueBeam linac, and 10,861 continuous EPID images were recorded (11.3 fps) and used for the DL training for a 6xFFF model. The transfer-learning model for a second linac was trained using a calibration plan to evaluate the efficiency in PDNet model adjustability by optimizing the last layer with core module parameters frozen. For validation, the PDNet models were tested with a new patient plan, where the prediction was compared with EPID measurements per field for the absolute dosimetry using gamma analysis (2%/2mm criteria) and that of a commercial solution.
Results: When using our deep learning model, PDNet achieved a higher gamma passing rate for all tested fields compared to the commercial solution. We show that the linac-I trained PDNet model achieved an overall passing rate of 98.7%, 97% on linac-I and linac-II, compared to 93.2% and 83.8% of the commercial dosimetry results, respectively. With less than two hours of transfer learning, the model improved the passing rate to 98.0% on linac-II.
Conclusion: We have developed a novel DL-based EPID dosimetry algorithm and demonstrated the clinical feasibility with high efficiency in adaptation/recalibration to other machines.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email