Research On Dynamic PET Image Reconstruction Based On The Kernel Method And Superpixel Clustering

Positron emission tomography(PET),a powerful functional imaging modality,can measure the spatial distribution of a radiotracer in vivo.Dynamic PET scan generally results in multiple time frames ranging from seconds to minutes in a series starting from the moment of tracer injection.Utilizing these multi-frame dynamic activity images,one can perform tracer kinetic modeling at the voxel or region of interest(ROI)level to obtain important physiological or biochemical parameters.However,accurate tracer kinetic modeling requires very short time frames,which result in low statistic counts and high noise in each frame.It is always a challenge to improve the quality of dynamic PET images.There are two main ways to improve the quality of PET images.One is to improve the quality of reconstructed images by improving the reconstruction algorithm,and the other is to reduce the image noise by post-reconstruction processing.This paper focuses on improving the quality of dynamic PET images.1.We proposed a median nonlocal means(MNLM)-based kernel method for dynamic PET image reconstruction.The method models PET image intensity in each pixel as a function of a set of features,which is derived from median nonlocal means of pre-reconstructed composite images.Then the model was incorporated into the forward model of PET projection data,and the coefficients of each features were estimated by maximum likelihood(ML)method.Using simulated and clinical low-count dynamic data,the quantitative performance of the proposed MNLM kernel method was investigated and compared with the maximum-likelihood method,conventional kernel method,and nonlocal means kernel method.Experiment results show that the constructed MNLM kernel can effectively extract the structure prior information from composite images,which improves the reconstruction of individual frame and achieves better image edges for the low-count dynamic PET.2.We propose a multiscale superpixel clustering-based dynamic PET reconstruction algorithm.This study aims to enhance the quantitative accuracy of individual dynamic frames via the introduction of priors based on multiscale superpixel clusters.The clusters are derived from pre-reconstruction composite images using superpixel clustering followed by fuzzy c-means(FCM)clustering.A multiscale aggregation is exploited during the superpixel clustering to generate multiscale superpixel clusters.Then,maximum a posteriori(MAP)PET reconstruction with different-scale clusters is separately applied to individual frame and fused to generate the final result.Using simulated and clinical dynamic PET data,the quantitative performance of the proposed method is compared with the maximum-likelihood expectation-maximization(MLEM),Bowsher method,and kernel method.Experiment results show that the constructed cluster prior can preserve the activities of the tumor region better while having substantially less background noise compared to other methods.3.We propose to apply the self-supervised Self2Self image restoration framework to dynamic PET image denoising.We use the Bernoulli sampled image of the static image rather than the noisy image as the network input.Thus,this method can not only utilize the structure prior information from the static image but also keep the advantages of Self2Self.This method does not require pre-training and does not require high-quality PET images.Simulation and clinical experiment results show that this method can achieve a better denoising effect and preservation of image structure than the traditional filter method and conditional deep image prior method.