Super-resolution in PET images using space variant kernels

Objectives Positron emission tomography (PET) images suffer from limited spatial resolution. In order to overcome this problem, several super-resolution (SR) algorithms were suggested. However, the existing PET image SR algorithms do not consider the problem of space variant resolution. In this paper, we present an image-domain SR scheme by applying space variant blur kernels.

Methods SR approach synthesizes a high-resolution (HR) image by using multiple low-resolution (LR) images. To obtain the relationship between HR image and LR images for SR, we need a point spread function (PSF) at each HR image pixel position. To obtain PSFs, we first acquire a sinogram of a point source at every pixel position. Using a sinogram, we then reconstruct an image to obtain a PSF or a blur kernel at the corresponding position. We use the ordered subsets expectation maximization (OSEM) algorithm for image reconstruction. A HR image is synthesized utilizing those space variant blur kernels in the SR process.

Results We evaluate the performance of the proposed method by using a mathematical phantom. We obtain point source sinograms by modeling a microPET R4 system. The modeling simulates the radial blurring effect depending on the point source location in the system. Using an original HR image, we generate four sub-pixel shifted LR images by using the obtained PSF kernels. We then synthesize a HR image. Comparing the synthesized HR image with the original one, we can note that the proposed SR algorithm can improve the resolution especially in the off-center compared to the existing SR algorithm which uses a space invariant box-shaped kernel.

Conclusions Introducing space variant blur kernels to image domain SR, we attempt to improve the PET image resolution. The simulation results show that the proposed SR algorithm can deliver more improved spatial resolution than the existing SR algorithm. We will evaluate this algorithm by using a Monte Carlo simulation package GATE later on.

Research Support The research was supported by the Converging Research Center Program through the Ministry of Education, Science and Technology (2010K001093)