Dose Characteristics of a Deep Learning Model for EPID-based

Background: Images from onboard electronic portal imaging devices (EPID) contain dose information that can be converted into dose maps. A cycle-consistent generative adversarial network (CycleGAN)-based model was developed for two-dimensional (2D) EPID dosimetry, and the dose characteristics of the model were evaluated carefully. Materials and Methods: All the measurements were done on a linac equipped with an EPID detector. This experiment involved: (1) assessing the dose characteristics of the EPID and (2) using a commercial treatment planning system to calculate dose distributions in a slab phantom, which were taken as the ground truth in CycleGAN-based models for converting the EPID images to 2D dose maps. There were about 780 beams delivered to EPID through a slab phantom. There were two normalization methods (NM): I: based on the highest possible value: 65,535 (16 bit); II: by its own maximum pixel value. To evaluate the model, gamma analyses between the ground truth and the output were performed with in-house software; and the dose linearity of the model was checked carefully. A comparative analysis was conducted to evaluate the outcomes stemming from two distinct NMs applied to the input data. Results: The dose characteristics of the EPID demonstrated exceptional precision. Notably, the beam output factors exhibited considerable variations with the increasing thickness of the phantom. Specifically, when the phantom thickness surpassed 12 cm, the trend lines exhibited a pronounced linearity. Deep learning models efficiently transformed EPID images into planar dose maps, albeit exhibiting dose nonlinearity that could be mitigated by choose the suitable normalization medthods. The mean pass rates of gamma analyses (3 mm, 3%) of data normalized by the way I or II were 85.5%, 97.9%, respectively. Conclusion: EPID is an excellent flat-panel detector that captures images rich in dose information, which can be effectively transformed into precise planar dose maps using CycleGAN-based models. The trained model could be used in the quality assurance of treatment plans.