A GPU-accelerated fully 3D OSEM image reconstruction for a high-resolution small animal PET scanner using dual-ended readout detectors

Abstract

In this work, a GPU-accelerated fully 3D ordered-subset expectation maximization (OSEM) image reconstruction with point spread function (PSF) modeling was developed for a small animal PET scanner with a long axial field of view (FOV). Dual-ended readout detectors that provided high depth of interaction (DOI) resolution were used for the small animal PET scanner to simultaneously achieve uniform high spatial resolution and high sensitivity. First, we developed a novel sinogram generation method, in which the dimension of the sinogram was determined first and then an event was assigned to a few neighboring sinogram elements by using weights that are inversely proportional to the distance from the measured line of response (LOR) to the LOR of the sinogram elements. System geometric symmetry, precomputation of LOR-driven ray-tracing and texture memory were applied to accelerate the GPU-based reconstruction. We developed a spatially variant PSF model where the PSF parameters were obtained by using point source images measured at 18 positions in the FOV and a spatial invariant PSF model where the PSF parameters were obtained by using only one image measured at the center FOV. The performance of the image reconstruction method was evaluated by using simulated phantom data as well as phantom and in-vivo mouse data acquired on the scanner. The results showed that the proposed reconstruction method provided better spatial resolution, a higher contrast recovery coefficient and lower noise than the OSEM reconstruction and was more than 1000 times faster than the CPU-based reconstruction. The spatially variant PSF model did not result in any spatial resolution improvement compared to the spatial invariant PSF model, and thus, the latter that is much easier to implement in image reconstruction and can be used in a small animal PET scanner using detectors with very high DOI resolution. A whole body 18F-FDG mouse image with high resolution and a high contrast to noise ratio was obtained by using the proposed reconstruction method.

Design and system evaluation of a dual-panel portable PET (DP-PET)

“…The output panelgram had a dimension of 35 × 56 × 35 × 56, with a total of 3,841,600 elements, which occupied approximately 29 MB of computer memory. A GPU-accelerated iterative image reconstruction method was developed based on the maximum likelihood expectation maximization (MLEM) algorithm and point spread function (PSF) resolution modeling [ 28 ]. The GPU kernel function was implemented to compute the forward projection and back projection in the line-of-response (LOR) units.…”

“…The RM was spatially variant and was applied in the reconstruction program by a PSF operation before projection as well as a transposed PSF operation after back projection. According to the results of previous experiments, including RM in the reconstruction increased the calculation time by approximately 10% [ 28 ]. The scale of the output reconstructed image was 140 × 224 × 160, with a voxel size of 0.713 × 0.713 × 1.0 mm 3 .…”

Design and System Evaluation of a Dual-Panel Portable PET (DP-PET)

“…After connecting with the cold source, it can quickly release heat through thermal silica and be reused. resolution modelling [28]. The GPU kernel function was implemented to compute the forward projection and back projection in the Line-of-response (LOR) units.…”

LCPR-Net: low-count PET image reconstruction using the domain transform and cycle-consistent generative adversarial networks

“…Typical iterative reconstruction algorithms include the maximum-likelihood expectationmaximization (MLEM) method (8), the improved version of the MLEM method (9) and the ordered subset expectation maximization (OSEM) method (10). Iterative reconstruction methods (11) usually optimize the objective functions that combine noise models in the sinogram domain and the image domain or prior knowledge in the image domain. According to the different prior knowledge sets, priors can be divided into nonlocal means (NLM) priors (12), dictionary learning priors (13)(14)(15)(16), and total variation (TV) priors (17)(18)(19)(20).…”