Exploring Deep Fully Connected Residual Neural Networks for Initial Energy Estimations of Compton Camera Based Prompt Gamma Imaging Data for Proton Radiotherapy

Abstract

Proton radiotherapy uses a beam of protons to irradiate cancer tissue. It has been suggested that real-time imaging can be used to help optimize treatment delivery via prompt gamma ray image data collected with a Compton camera. When a prompt gamma deposits energy twice, it is called a “double” and physicists assume that it has deposited all of its energy after the second collision, thereby being absorbed and the total of the energy depositions is considered to be the initial energy of the prompt gamma ray. This initial energy is used during reconstruction to help determine the origin of the prompt gamma and leads to a well formed reconstruction but the assumption that a double deposits all of its energy is not always true leading to improper origins based on incorrect initial energies. Here, we present a deep residual fully connected regression neural network which can make estimations of the initial energy of double events on data generated by a Monte Carlo simulation using Compton camera detector effects. We conduct a hyperparameter search to explore different models. We then present and discuss the results of the currently best performing regression neural network model. We suggest further improvements that can help further reduce loss and improve model estimations for reconstructed images.