Ziel/Aim:
In quantitative PET/CT or when used in computer-aided diagnosis, textural feature
analysis must base on reproducible values. This work investigated exposure dependency
of second-order statistical textural features derived from grey level co-occurrence
matrices (GLCMs). Exposure E was defined as phantom activity concentration times acquisition
duration.
Methodik/Methods:
A homogeneous cylindrical Ge-68 phantom with 9.6 kBq/ml activity concentration was
imaged on a Siemens Biograph mCT with acquisitions ranging from 3 s to 10861 s. Images
with differing isometric voxel sizes were reconstructed with filtered back-projection
(FBP), ordered subset expectation maximization (OSEM) and the Siemens TrueX algorithm.
Eleven GLCM derived features – taken from a 50 mm wide cube at the centre of the phantom
– were calculated and plotted as functions of exposure f(E). Feature stability was
defined for df/dE -> 0.
Ergebnisse/Results:
Feature values from FBP reconstructions with 4 mm sized voxels were the most stable,
whereas values from TrueX reconstructions with 1.5 mm sized voxels varied up to three
orders of magnitude across different exposures. Some features shared common islands
of stability showing exposure invariant metrics. Stability in ten features was reached
in 4 mm FBP and OSEM acquisitions after around ½ min; 1.5 mm TrueX acquisitions took
more than 30 min to achieve stability in nine features.
Schlussfolgerungen/Conclusions:
Textural features derived from GLCMs vary strongly with exposure, but islands of stability
exist for some. This exposure invariance results in comparable and reproducible feature
values, provided that any such stable regions are determined prior to the actual PET/CT
measurements. Adjusting exposures to expected activity concentrations will prevent
image count statistics from dominating texture values over true object inhomogeneity.
Only within stable regions of exposure will textural features analysis make sense
in PET/CT.