A log likelihood based method for recovery of localized defects in PET attenuation-correction images

Abstract

In PET/CT, CT data is used to produce a PET attenuation correction. Contrast administered to enhance CT images can confound algorithms that scale CT data to a 511 keV attenuation image. The result may contain localized regions of artifactually high values. We present a method to identify artifacts by using the PET emission data. To start, suspect regions of an attenuation map are defined. Such a region might be, for example, all pixels with high attenuation values under the assumption that such values are either correct or the result of improper contrast scaling. We assume that incorrectly assigned attenuation should have a known alternate value, for example, that of blood. The algorithm visits each pixel in the defined region and provisionally assigns it one of the two possible attenuation values. It adjusts the activity value of that pixel to maximize the emission log likelihood. This is repeated for the alternate attenuation value. The pixel is assigned the pair of activity and emission values that maximizes the log likelihood. After all pixels have been visited an emission update of the full image is performed. These steps can be iterated. In a digital anthropomorphic phantom study the suspect region comprised 26 artifactuality high valued attenuation pixels (contrast) and 603 truly high pixels (bone) in a slice at chest level. The algorithm was tested in the presence of statistical noise spanning the range from 1000 to 10,000,000 nominal sinogram counts with eight statistical variates at each count level. Results show that the algorithm works well. Remaining or induced attenuation artifacts, when carried over to the reconstructed emission image, were generally found to be insignificant compared to image noise