Scatter correction in maximum-likelihood reconstruction of PET data

To obtain quantitative PET (positron emission tomography) images with the ML (maximum likelihood) reconstruction algorithm, the authors investigated the inclusion of a correction for scatter. They implemented the spatially variant convolution method of M. Bergstrom et al. (J. Comput. Assist. Tomogr., vol.7, p.42-50, 1983) which assumes that scatter is independent of depth and collapses the problem to a projection-by-projection scatter model. The model was implemented in three ways: subtraction of scatter estimated from measured projections prior to reconstruction; inclusion of a scatter estimate from the measured projection data in the iteration loop; and inclusion of a scatter estimate in the iteration loop, based on the previous iteration´s estimate of trues from the image. The reconstructions were performed on an Intel iPSC/860 hypercube computer. Analysis of the convergence, bias, and noise properties of the three methods of scatter correction demonstrated only slight differences between the methods for real phantom data taken on the Scanditronix PC2048-15B brain PET scanner. The structure of this ML algorithm permits direct extension to a more comprehensive model of scatter