Dual-tracer PET Image Reconstruction Based On Convolutional Network

Ideal multi-tracer Positron Emission Tomography(PET)technology is to inject multiple tracers simultaneously with the ability to observe the physiological activities of tissue in various aspects.This definitely provides more comprehensive understanding of diseases,but also reduces the scanning time and cost for patients,or improves the use efficiency of PET scanner.Due to the limitation of the PET imaging physics,the signals emitted by different tracers are the same(all are 511keV photon signals).As a result,although many clinical trials have confirmed the significance of multi-tracer PET technology,this technology has not been applied in clinical.This thesis tries to address the dual-tracer PET image reconstruction problem from standard dose and low dose injection respectively.The main content and contributions of this thesis are as follows:1.Faced with the problem of noise difference of each tracer due to different concentration distributions,a three-dimensional convolutional encoder-decoder network based on multi-task learning is proposed.The dual-tracer PET image reconstruction is viewed as two related tasks being performed simultaneously.In the proposed model,an encoder is shared between the two tasks,and the decoder is used separately.Considering the importance of time information in the reconstruction of dual-tracer PET images,we use a three-dimensional convolutional network to make full use of time and space information.The verification of simulation experiment and rat experiment proves the practicability and progressive of the method.2.Considering that the multi-tracer PET technology will increase the dose of the tracer,low-dose imaging is necessary.However,no related work has been published yet.Therefore,a model based on the attention mechanism is proposed to estimate the standard dose single tracer sinogram from the low-dose dual tracer sinogram,that is,to achieve the standard dose estimation and dual tracer signal separation at the same time.We cascaded the FBP-Net for reconstruction after the proposed model,and verified the effects of the proposed method in simulation experiments and rat experiment.The results of experiments confirmed the robustness and accuracy of the proposed method.