Automated liver reference region localization in PET/CT in presence of non-physiological tracer uptake

Introduction: The measurement of physiological tracer uptake within a reference region in the liver is required for standardized PET image assessment in several clinical applications. In order to streamline measurements and improve repeatability, semi-automated and automated algorithms have been developed for determining a reference volume of interest (VOI) in the liver. Nonetheless, positioning of the VOI in healthy liver tissue remains a challenge considering the presence of artifacts or non-physiological liver radiotracer uptake. We therefore developed and evaluated an automated algorithm for determining a suitable liver reference region based on combined information from PET and CT images.

Methods: The proposed method is aimed at determining the position of a 3 cm-diameter sphere contained in a liver region with uniform tracer uptake and attenuation characteristics, respectively in PET and CT images, within the sphere and its close proximity. First, a binary mask is determined by segmenting the liver based on the CT image using a previously described deep learning method (Yang et al. 2017). Secondly, the center of a 3 cm-diameter sphere contained in the liver mask is determined such that the image values within a concentric 4 cm-diameter sphere and the liver mask have minimal geometric mean between coefficients of variation of PET and CT values. The coefficients of variation of image values for all possible sphere positions are obtained from the PET/CT image through a combination of convolutions in the Fourier domain and unary voxel-wise operations allowing faster computation. The proposed method was evaluated in a set of PET/CT scans presenting foci with non-physiological tracer uptake either in the liver or with spill-over into the liver. The proposed method was compared to an automated method for positioning the liver VOI based solely on the CT image using an ensemble machine learning landmark detection algorithm (Tao et al. 2011). The reference for comparison was the frequency with which the determined VOI intersected non-physiological uptake regions annotated by an experienced nuclear medicine physician.

Results: PET/CT scans of 130 subjects were included in the analysis (¹⁸F-FDG: 122, ⁶⁸Ga-PSMA-11: 8). In one subject, both the proposed method and the landmarking algorithm failed, not detecting respectively the liver mask and landmark. In the remaining 129 subjects, the resulting liver VOI intersected non-physiological tracer uptake regions for the proposed method in 4 subjects (3%), for the landmarking algorithm in 13 subjects (10%), for both methods in 3 subjects, and for neither method in 115 subjects. A significant difference in frequency of non-physiological uptake intersection with the reference liver VOI was found for the two methods (significance level set at 5%, McNemar’s exact test p=0.01).

Conclusions: The proposed automated liver reference VOI algorithm based on combined information from PET and CT images performed better at placing VOI’s which avoid non-physiological tracer uptake compared to a reference method based on CT only. The proposed fully automated method holds promise to enable robust and reproducible measurement of physiological tracer uptake in the liver for standardized PET image assessment according to established criteria.

• Download figure
• Open in new tab
• Download powerpoint