Predicting Linac Beam Properties From Sparse Measurements Using a Neural Representation Network for Fast Commissioning and Quality Assurance
Presentations
SU-H430-IePD-F4-3 (Sunday, 7/10/2022) 4:30 PM - 5:00 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 4
Purpose: To investigate the use of coordinate-based neural representation network and single vendor-provided beam dataset for linear accelerator (Linac) beam modeling and evaluate the beam model in predicting clinical beam data from limited measurements to facilitate commissioning and quality assurance (QA) procedures.
Methods: A sinusoidal neural representation network that consists of multilayer perceptron and periodic activation function was used to learn a mapping between 2D coordinates (1D spatial locations x field sizes) during beam scanning and the corresponding percentage dose. The network was trained using a vendor-provided 6 MV reference beam dataset and evaluated using clinical 6 MV beam datasets consist of 79 lateral beam profiles and 57 percentage depth dose (PDD) curves collected from multiple clinics. Only beam data with 10x10 cm² field size was assumed available and used to update the network weights. The network was then used to predict beam data with various field sizes ranging from 3x3 cm² to 40x40 cm². The predicted beam data was compared to the ground truth by calculating both the mean absolute error and the 1D gamma passing rate.
Results: For lateral beam profile prediction, the averaged mean absolute error between predicted and ground truth percentage dose was 0.9%. Averaged gamma passing rates with pass criteria of 1%/1mm and 2%/1mm were 96% and 98% respectively. The maximum improvement in gamma passing rate using model-predicted data instead of vendor-provided reference was 19% (1%/1mm) and 20% (2%/1mm). Similar results were obtained for PDD prediction with an averaged mean absolute error of 0.3% and an averaged gamma passing rate (1%/1mm) of 98%.
Conclusion: Neural representation network can model Linac beam properties and predict clinical beam data accurately from sparse beam measurements. The proposed beam model has the potential of simplifying the clinical beam scanning process with much reduced beam data collection.
Keywords
Quality Assurance, Linear Accelerator, Commissioning
Taxonomy
TH- External Beam- Photons: Quality Assurance - Linear accelerator
Contact Email