Parameterization of a model-based 3D whole-body PET scatter correction

Parameterization of a fast implementation of the Ollinger model-based 3D scatter correction method for PET has been evaluated using measured phantom data from a GE PET Advance/sup TM/. The Ollinger method explicitly estimates the 3D single-scatter distribution using measured emission and transmission data and then estimates the multiple-scatter as a convolution of the single-scatter. The main algorithm difference from that implemented by Ollinger (1996) is that the scatter correction does not explicitly compute scatter for azimuthal angles; rather, it determines 2D scatter estimates for data within 2D "super-slices" using as input data from the 3D direct-plane (non-oblique) slices. These axial super-slice data are composed of data within a parameterized distance from the center of the super-slice. Such a model-based method can be parameterized, choice of which may significantly change the behavior of the algorithm. Parameters studied in this work included transaxial image downsampling, number of detectors to calculate scatter to, multiples kernel width and magnitude, number and thickness of super-slices and number of iterations. Measured phantom data included imaging of the NEMA NU-2001 image quality phantom, the IQ phantom with 2 cm extra water-equivalent tissue strapped around its circumference and an attenuation phantom (20 cm uniform cylinder with bone, water and air inserts) with two 8 cm diameter water-filled non-radioactive arms placed by it's side. For the IQ phantom data, a subset of NEMA NU-2001 measures were used to determine the contrast-to-noise, lung residual bias and background variability. For the attenuation phantom, ROIs were drawn on the nonradioactive compartments and on the background. These ROIs were analyzed for inter and intra-slice variation, background bias and compartment-to-background ratio. Results: In most cases, the algorithm was most sensitive to multiple-scatter parameterization and least sensitive to transaxial downsampling. The algorithm showed convergence by the second iteration for the metrics used in this study. Also, the range of the magnitude of change in the metrics analyzed was small over all changes in parameterization. Further work to extend these results to other more realistic phantom and clinical datasets is warranted.