Single-Scan Dual-Tracer Separation Network Based on Pre-trained GRU

In this paper, a novel network based on gated recurrent unit (GRU) is proposed for separating single-scan dual-tracer PET mixed images. Compared to conventional methods, this method can separate dual-tracer that are simultaneously injected or even labeled with the same marker, and do not require arterial blood input function. The proposed 4-layer network denoises the time activity curves (TACs) extracted from the dynamic dual-tracer reconstruction images with noise by pre-training the parameters in the first and second layer, and then uses TAC time information for dual-tracer separation. During the training stage, we optimize the network by minimizing the mean square error (MSE) objective function of the separated predicted value and ground truth. Monte Carlo is used to simulate the PET sampling environment with the mixed dual-tracer \({^{62}}\)Cu-ATSM+\({^{62}}\)Cu-PTSM and \({^{18}}\)F-FDG+\({^{11}}\)C-MET. Calculating the bias and variance to quantitatively analyze the results, we demonstrate that this method is more robust and better separation than the similar methods.