Scatter Correction Requirements for Likelihood-Based Attenuation Artifact Correction in PET

Artifacts have been observed in CT-based, 511-keV attenuation images for PET attenuation correction. Artifact detection and correction algorithms are being developed based on the Poisson log likelihood function (LLF). The LLF quantifies the level of consistency between an activity and attenuation image pair and the measured emission sinogram. We have shown that the LLF can discriminate more-artifactual from less-artifactual attenuation images at clinical noise levels. Here we use a digital anthropomorphic phantom to examine the effect of scatter on LLF performance. Base sinograms were formed by forward projecting a true activity distribution through a true attenuation distribution. Scatter was simulated by filtering the base sinograms. Noise was considered by producing 128 Poisson-randomized variates of the filtered base sinograms for each of 7 assumed count levels. An artifactual attenuation image was produced by modifying mu-values of regions of the true attenuation image. For each sinogram, two OSEM reconstructions were performed and two LLF values obtained using the artifactual and true attenuation. Accuracy was calculated as the fraction of cases in which the LLF preferred the artifact-free attenuation image. Additional studies were performed to test several scatter correction procedures. Findings show that scatter may render the LLF unreliable even in the limit of high-count scans but that its usefulness is restored by applying an appropriate correction.