Direct mapping from PET coincidence data to proton-dose and positron activity using a deep learning approach

Objective. Obtaining the intrinsic dose distributions in particle therapy is a challenging problem that needs to be addressed by imaging algorithms to take advantage of secondary particle detectors. In this work, we investigate the utility of deep learning methods for achieving direct mapping from detector data to the intrinsic dose distribution. Approach. We performed Monte Carlo simulations using GATE/Geant4 10.4 simulation toolkits to generate a dataset using human CT phantom irradiated with high-energy protons and imaged with compact in-beam PET for realistic beam delivery in a single-fraction (∼2 Gy). We developed a neural network model based on conditional generative adversarial networks to generate dose maps conditioned on coincidence distributions in the detector. The model performance is evaluated by the mean relative error, absolute dose fraction difference, and shift in Bragg peak position. Main results. The relative deviation in the dose and range of the distributions predicted by the model from the true values for mono-energetic irradiation between 50 and 122 MeV lie within 1% and 2%, respectively. This was achieved using 10⁵ coincidences acquired five minutes after irradiation. The relative deviation in the dose and range for spread-out Bragg peak distributions were within 1% and 2.6% uncertainties, respectively. Significance. An important aspect of this study is the demonstration of a method for direct mapping from detector counts to dose domain using the low count data of compact detectors suited for practical implementation in particle therapy. Including additional prior information in the future can further expand the scope of our model and also extend its application to other areas of medical imaging.