Using Deep Learning to Enhance Compton Camera Based Prompt Gamma Image Reconstruction Data for Proton Radiotherapy

Abstract

Proton beam radiotherapy is a cancer treatment method that uses proton beams to irradiate cancerous tissue while simultaneously sparing doses to healthy tissue. In order to optimize radiational doses to the tumor and ensure that healthy tissue is spared, many researchers have suggested verifying the treatment delivery through real-time imaging. One promising method of real-time imaging is through a Compton camera, which can image prompt gamma rays emitted along the beam's path through the patient. However, the images reconstructed with modern reconstruction algorithms are often noisy and unusable for verifying proton treatment delivery due to limitations with the camera. This paper demonstrates the ability of deep learning for removing false prompt gamma couplings and correcting the improperly ordered gamma interactions within the data for the case of Triples and Doubles-to-Triple events.

Deep Learning for Classification of Compton Camera Data in the Reconstruction of Proton Beams in Cancer Treatment

“…Our results show that it is in fact feasible to apply a fully connected neural network to this problem. Further tests as well as discussions of the effect classification has on actual reconstructions are provided in [3].…”

“…We propose that a neural network can be trained to determine if a given event is "false" or, if true, the correct camera interaction ordering. A more complete discussion is available in [3].…”

Deep residual fully connected neural network classification of Compton camera based prompt gamma imaging for proton radiotherapy

“…Our initial studies in [28] showed that our general NN configuration is capable of identifying both true and false triple scatters. In this study, we have expanded the complexity of the Frontiers in Physics frontiersin.org network to now consist of 64 residual blocks with eight fully connected layers per block yielding a total of 512 hidden layers.…”

“…This more complex network, as shown in these results, is able to handle more complex triple scatter datasets, that include true and false events, as well as DotT and false triples. Similar to [28], all data used for training and validation are simulated and are produced by the MCDE model discussed in Section 2.1.1. We use an 80/ 20 training/validation data split as is suggested in [15] but since our data is entirely simulated this split is arbitrary and done for consistency with other research within the domain.…”

Machine learning-based event recognition in SiFi Compton camera imaging for proton therapy monitoring

“…These problems make the reconstructed PG depth profile more noisy and unusable in clinical use (Polf and Parodi 2015). Therefore, several research groups have proposed machine learning-based event selection as a worthwhile approach to reduce background in Compton camera imaging (Barajas et al 2021, Muñoz et al 2021, Polf et al 2022 and deliver more precise doses to the tumors during proton therapy (Gueth et al 2013, Basalyga et al 2020. To evaluate the performance of our Compton camera prototype currently under investigation, a machine learning software based on the ROOT (Brun and Rademakers 1997) toolkit for multivariate data analysis 'TMVA' (Hoecker et al 2007) is proposed to properly discriminate Compton events from background events.…”