Investigating the Impact of Susceptibility Artifacts on Adjacent Tumors in PET/MRI through Simulated Tomography Experiments

Abstract

For quantitative PET imaging, attenuation correction (AC) is mandatory. Currently, all main vendors of hybrid PET/MRI systems apply a segmentation-based approach to compute a Dixon AC-map based on fat and water images derived from in- and opposed-phase MR-images. Changes in magnetic susceptibility pose major problems for MRI, which may lead to artifacts resulting in tissue misclassification in the segmented AC-map. Cases have been reported where the liver has been misidentified as lung tissue due to iron overload, e.g. from hemochromatosis or iron oxide MR contrast agents, resulting in severe underestimation of PET-quantification.

In this thesis, simulated tomography experiments were conducted to investigate the impact of susceptibility artifacts on adjacent tumors, focusing on the misclassification of liver tissue as lung tissue. A digital phantom was programmed, and synthetic tumors and artifacts were introduced into a realistic PET/MRI patient dataset. The data were reconstructed with attenuation maps both with and without artifacts to compute the relative error (RE) in tumor uptake.

It was shown that relevant errors can be introduced to tumors adjacent to the artifact. A strong inverse square relationship between the distance (d) of the center points of a tumor and an artifact was found with the RE. Further, because the RE was known to be proportional to the volume (V) of misclassified tissue, it was shown that it is possible to obtain a linear equation describing the RE using only V and d. However, this assumes similar information, i.e activity and attenuation, along the common line of responses (LORs) of the artifact and tumor.

A correction method was developed to correct for lung-liver misclassifications. The proposed method uses the already acquired opposed-phase Dixon images, which are less sensitive to susceptibility changes. It successfully corrected 96% of misclassified tissue down to a 50% MR-signal reduction from the liver. The method benefits from using already acquired data to correct the artifacts, and may be made fully automatic to function in real-time.