Statistical Image Reconstruction: 2D PET Results

PET Phantom, GE Advance System

FBP .......... PL

These images were reconstructed from a PET scan of a thorax phantom with a hot liver in a warm background, acquired by the University of Washington PET group. Slice 30 of 35 (counting from 0) is displayed, which had 8.6M counts in the emission sinogram.

The left image was reconstructed using FBP with this PET system's conventional attenuation correction factors (from a 25M count measured transmission scan).

The right image was reconstructed by the following statistical approach. I first applied the grouped coordinate ascent algorithm (IEEE T-MI, Apr. 1997) to compute a penalized-likelihood estimate of the attenuation map using a nonquadratic penalty function. Then I reprojected that attenuation map estimate to form attenuation correction factors. These factors, along with sinogram normalization factors, were incorporated into the system model for the space-alternating generalized EM (SAGE) algorithm which was used to reconstruct the above image using penalized-likelihood (PL) with a quadratic penalty function. Since the sinogram had been precorrected for random coincidences, the shifted-Poisson method of Yavuz and Fessler was applied with the SAGE algorithm.

For display, the negative values in the FBP image were set to zero (black). There are no negative values in the PL image since nonnegativity is enforced by the SAGE algorithm.

Here are links to papers describing the algorithms.