Improving Proton Beam Radiotherapy by Classifying Simulated Patient Data in Compton Camera Imaging with Neural Networks

Abstract

Proton beam radiotherapy is an advanced cancer treatment utilizing high-energy protons to destroy tumor matter. When a proton beam interacts with a patient’s body, it emits prompt gamma rays, which are detected by a Compton camera. However, image reconstruction of the beam path from these scatterings is often unusable due to mischaracterized scattering sequences and excessive image noise. To address this, machine learning models were developed to classify the scattering events. Multiple novel robust-volume datasets simulating particle interactions with human tissue were generated using Duke University CT scans and Geant4 and Monte-Carlo Detector Effects (MCDE) software. Novel implementations of a Event Classifier Transformer and a 1D Convolutional Neural Network (CNN) were developed to better address spatial scattering relationships. The prior models, along with a Fully-Connected Neural Networks (FCN) and Long Short-Term Memory Neural Network (LSTM), were optimized through large-scale hyperparameter studies using a new automated tuning framework built into the Big-Data REU Integrated Development and Experimentation (BRIDE) machine learning pipeline. From hyperparameter tuning and larger datasets, FCN and LSTM models achieved significant 17-18% numerical accuracy increases from any previous work. With minimal overfitting, these models offer much greater generalizability. Transformer and CNN models were not as well suited to patient data but still achieved accuracies comparable or greater than previous work.