EPIDEEP: Using a Deep Learning Model to Predict In Vivo Electronic Portal Imaging Device (EPID) Transit Images

Abstract

To create a deep-learning model to predict in-vivo electronic portal imaging device (EPID) transit images for IMRT treatments. This model was created to predict in-vivo images to identify machine and patient-related errors that occur during beam delivery and are undetectable with current QA approaches. The deep-learning model can make image predictions much faster than Monte Carlo approaches, making image prediction feasible for application in online adaptive radiotherapy. Additionally, the model does not rely on any proprietary information and can be easily utilized by other clinics. Our approach separates the primary and scatter components of in-vivo transit images. The attenuation of primary radiation reaching the EPID panel is modeled analytically, using attenuation measurements from phantoms of known thicknesses. The scatter component is estimated using a convolutional neural network (CNN). The CNN training uses information from the on-treatment cone-beam CTs (CBCTs), and a pretreatment EPID image with no patient in the beam. We acquired 193 IMRT fields/images from 118 patients previously treated on the Varian Halcyon. Treatment sites included the pelvis, abdomen, lungs, and extremities. CBCTs were collected immediately before treatment, to provide an accurate representation of the anatomy. A 3-channel input image was used, consisting of the pretreatment EPID image, a ray tracing projection through the CBCT to the EPID panel, and a projection to isocenter. Model training:validation:test set ratios were 133:20:40 images. The primary and scatter components are added together to give the predicted transit image. Prediction accuracy was assessed by comparing model-predicted and measured in-vivo EPID images with a 3%/3mm and 5%/3mm gamma pass rate. The gamma pass rate for the patients in the training:validation:test was 91.5%:90.4%:92.1% for 3%/3mm and 96.7%:96.6%:97.0% for 5%/3mm. The model can make image predictions in 20 milliseconds. The poor passing rates of some images may be due to CBCT artifacts and patient motion that occurs between the time of CBCT and treatment. This model can predict in-vivo EPID images with an average gamma pass rate greater than 90%. Image predictions from this model can be used to detect in-vivo treatment errors and changes in patient anatomy, providing an additional layer of patient-specific quality assurance. The speed of image predictions is 20 milliseconds, making use feasible for online adaptive treatments, which currently do not utilize patient-specific measurements of the delivered radiation. Upcoming studies will assess the model's ability in detecting clinically relevant errors and changes in patient anatomy that can adversely affect treatment. Future goals include acquiring more data to further improve the model and extending the model to make predictions for VMAT treatments.