Improvements to the Deep Learning Classification of Compton Camera Based Prompt Gamma Imaging for Proton Radiotherapy

Author/Creator ORCID

Date

2020-05

Department

Program

Citation of Original Publication

Basalyga, Jonathan N.; Barajas, Carlos A.; Kroiz, Gerson C.; Gobbert, Matthias K.; Maggi, Paul; Polf, Jerimy; Improvements to the Deep Learning Classification of Compton Camera Based Prompt Gamma Imaging for Proton Radiotherapy; Technical Report HPCF–2020–29; http://hpcf-files.umbc.edu/research/papers/BasalygaBarajas_Summer2020.pdf

Rights

This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.

Abstract

Real-time imaging has potential to greatly increase the effectiveness of proton beam therapy for cancer treatment. One promising method of real-time imaging is the use of a Compton camera to detect prompt gamma rays, which are emitted along the path of the beam, in order to reconstruct their origin. However, because of limitations in the Compton camera’s ability to detect prompt gammas, the data are often ambiguous, making reconstructions based on them unusable for practical purposes. Deep learning’s ability to detect subtleties in data that traditional models do not use make it one possible candidate for the improvement of classification of Compton camera data. The base network can be made cheaper via reducing hidden layer count while maintaining comparable classification performance. Additionally, even a simple training schedule can show improvements in the training process. Several variations of the network showed promise in their ability to classify multiple beam energies. However more improvements need to be made to the network for the performance on multiple beam energies to meet our goal of 90% classification accuracy.