Separation of dual-tracer PET signals using a deep stacking network

In this study, a method based on a deep stacking network is proposed to solve the signal separation problem of dynamic dual-tracer PET. The advantage of this method is that it avoids requirements for prior information of tracers, and a staggered injection. The proposed model is pre-trained with restricted Boltzmann machines and fine-tuned in a manner, which the output of the last training epoch was used as additional input in the current epoch to update the model parameters. We train the network to learn the complex relationship between dual-tracer time-activity curves and separated single tracer data using a mean square error objective function. Monte Carlo simulations are employed to test the accuracy and robustness of the proposed method on the total counts and reconstruction algorithm. Quantification results show that the proposed method outperforms the existing approach. Experiments with real data further validate previous results on synthetic data.