Dual-Tracer PET Imaging Signal Separation Based On Neural Network

Positron Emission Tomography(PET)is a highly sensitive and specific quantitative molecular imaging technology that can reveal the distribution of tracers in organisms and help early diagnosis of tumors and neurological diseases.Compared with traditional static single-tracer PET images,dynamic dual-tracer images can provide more complete disease information,improve diagnosis accuracy and reduce the scan times.However,there are still many challenges for dynamic dual-tracer PET imaging.Most traditional methods can only separate dual tracers injected at intervals,rely on kinetic models and arterial input function(AIF),and are sensitive to the types of tracers.In response to the above problems,this paper proposed a Mask-Based Bidirectional Gated Recurrent Unit(MB-BGRU)network framework,which can separate dual tracers injected at the same time and do not require AIF and dynamic model constraints.The network was divided into three parts:encoding module,separation module and decoding module.The BGRU layer can capture temporal information.Simulation experiments and real monkey brain experiments verified the robustness and feasibility of the network under shadow factors such as AIF,kinetic parameters,noise level,tracer types,template types,and sampling protocols.Further,in order to expand the clinical application of dual tracers,we proposed a model-based Block-Coordinate-Descent Encoder-Decoder(BCD-ED)neural network framework,which can reconstruct and separate the full-dose single-tracer images from the mixed dynamic dual-tracer sinogram.The network is divided into three modules.The reconstruction and denoising module combined the traditional iterative reconstruction algorithm with deep learning.The separation module used the encoder-decoder structure to learn the mapping relationship between the mixed concentration images and the single-tracer concentration images.Real monkey brain experiment verified the feasibility of the network under real conditions.