Range and dose verification in proton therapy using proton-induced positron emitters and recurrent neural networks (RNNs)

Online proton range/dose verification based on measurements of proton-induced positron emitters is a promising strategy for quality assurance in proton therapy. Because of the nonlinear correlation between the dose distribution and the activity distribution of positron emitters, we aim to establish their relationship using recurrent neural network models (LSTM, BiLSTM, GRU, BiGRU and Seq2seq). Simulations were carried out with a spot-scanning proton system using Geant4-10.3 toolkit and a CT-based patient phantom. The 1D distributions of positron emitters and radiation dose were obtained. Training data were modeled for different beam energy, irradiation positions and counting statistics. The prediction accuracy of range and dose were quantitatively studied. The impact of including anatomical information (HU values in CT images) on the prediction performance was investigated. The BiGRU demonstrates the most stable and accurate performance with good capability of generalization, especially with the inclusion of anatomical information. When the signal-to-noise ratio (SNR) of the 1D activity profiles is about 3, the range accuracy can be within 0.5 mm and the dose accuracy close to the peak region is  <5% (relative uncertainty between prediction and raw input for all datasets). The feasibility of proton range and dose verification using the RNN-based framework was demonstrated. The RNN-based framework promises to provide a reliable and effective way for online monitoring, quality assurance and ultimately allows for adaptive proton therapy.