Jerimy Polf, Department of Radiation Oncology, University of Maryland School of Medicine
Paul Maggi, Department of Radiation Oncology, University of Maryland School of Medicine
Jonathan Basalyga, Department of Mathematics and Statistics, UMBC
Gerson Kroiz, Department of Mathematics and Statistics, UMBC
Carlos A. Barajas, Department of Mathematics and Statistics, UMBC
Matthias K. Gobbert, Department of Mathematics and Statistics, UMBC
The advantage of proton beam therapy in cancer treatment is that the peak radiation dose is delivered at the end of the beam range, known as the Bragg peak (BP), with no dose delivered beyond. By using these characteristics of the BP, the radiation dose to the tumor can be maximized, with a greatly reduced radiation dose to the surrounding healthy tissue. During treatment delivery, protons in the beam can undergo nuclear interactions with atoms in the patient tissue, which can result in the instantaneous emission of secondary gamma rays called prompt gammas (PGs). Detection of the PGs, along with a suitable reconstruction method, can provide near real-time verification and assessment of the treatment delivery. However, such imaging requires fast detectors and rapid image reconstruction to be feasible.
Traditional reconstruction methods (SOE, MLEM) can take several minutes to produce an image, which would increase total patient treatment time and cannot be used for adaptive therapy. Additionally, the high count-rate environment reduces the detector’s data quality, further complicating reconstruction. Deep learning has the potential to provide guidance to the treatment, by being able to rebuild an image in seconds, even with degraded data. There has been research which use Convolutional Neural Networks to reconstruct everyday objects in 3D space, however in this paper we are attempting to reconstruct a 3D beam as seen inside the patient. We run the risk of the network becoming to computationally expensive during feature digestion and in response must be very careful in the configuration of a CNN which allows for multiple spatially dense 3D data input. This allows for the radiation to be monitored in near real time while the patient is being treated.