EPID image prediction using a U-net model.

Abstract: Radiation therapy is one of the most common treatment methods for cancer today, with an estimated half of patients requiring it at some point in the course of their illness. In radiation therapy, verification of treatment plans is an important process. Treatment verification can be performed by measuring the outgoing radiation from the treatment machine and comparing with the estimate from the treatment planning system (TPS), called the fluence map. One commonly used measuring device is an electronic portal imaging device (EPID), which can capture 2D digital images of the incoming radiation. The EPID does not image the fluence perfectly, and the images have some small but significant artifacts, caused by phenomena like scattering in the image detection plane. Before the fluence map and EPID images can be compared, some corrections need to be applied to either the fluence map, the EPID image, or both, in order to account for the aforementioned artifacts. In this thesis the use of a deep learning model to predict the EPID images from the fluence maps is investigated. The model should, given a fluence map, predict the resulting EPID image as accurately as possible. The model was trained using simple rectangular training fields and evaluated using fields from real, clinical treatment plans. The results show that the deep learning model’s predictions are an improvement over the fluence maps in terms of similarity to the real EPID image, indicating that the model has learned useful transformations to apply to the fluence maps. In terms of gamma evaluation pass rates, the pass rates increased by between 1:3 and 14:9 percentage points on the patient fields tested.