Dynamic PET Image Reconstruction Using the Wavelet Kernel Method

Abstract

Dynamic PET imaging is used to monitor the spatio-temporal distribution of a tracer in a tissue region. Dynamic PET can suffer from high noise; to address this problem, the kernel method has been developed for efficient dynamic PET image reconstruction. Previous kernel approaches used a Gaussian kernel to exploit nonlocal spatial correlations from image priors. The Gaussian kernel, has an undesired effect of smoothing high frequencies. In this work, we propose using a wavelet kernel with good energy compaction to further enhance kernel-based dynamic PET image reconstruction. The oscillation in the wavelet kernel can result in better representation of details in the final reconstructed images. We evaluated the wavelet kernel approach using patient data acquired from dynamic C-11 hydroxyephedrine (HED) PET imaging. Reconstruction results demonstrate that this wavelet kernel approach achieves better image quality than standard reconstruction and the Gaussian kernel approaches.