High Temporal Resolution Dynamic PET Image Reconstruction Based On Kernel Method

Dynamic PET imaging plays a key role in the field of observing the process of metabolism,exploitation of new medicine and parametric image reconstruction.But conventional dynamic PET imaging has a limited temporal resolution,and always suffers from pool signal to noise ration(SNR)due to low-count statistics from short time frames and the ill-posed nature of expectation-maximization(EM)based reconstruction algorithm.Kernel method in machine learning,when used in PET image reconstruction,has excellent ability to deal with the heavy noise from short frames,and the spatial-temporal kernel method proposed lately could also generate comparable dynamic PET images from shorter frame.In this paper,we implemented the spatial-temporal kernel method and applied it to the data collected from the 2 m PET/CT scanner(u EXPLORER)to generate 0.1 s ultra-high temporal resolution dynamic PET images and succeeded in extracting noiseless cardiac motion signal.We proposed a sparse constraint by a neighborhood and k nearest neighbors to compress the size of spatial kernel matrix without degrading the performance of kernel method.We also simulated two short range scanners by using partial of the total eight units of u EXPLORER,and applied the kernel method to them.The result demonstrated that the quality of short range scanner prior images may influence the performance of kernel method.In order to manage that,we proposed training a V-Net network with the short prior images as input and treating the prior images from long range scanner as ground truth to generate prior images with comparable quality as long range scanner.And we realized 1 s frame reconstruction in short range scanner under the guidance of improved prior images.