- https://doi.org/10.1029/2018JB017083
- doi:10.1029/2018JB017083
Abstract — We present a method for accurately measuring small, discrete features near the resolution limit of X-ray computed tomography (CT) data volumes with the aim of providing consistent answers across instruments and data resolutions. The appearances of small features are greatly impacted by the partial volume effect and blurring due to the point-spread function (PSF) of the data, and we call our approach the PVE method. Features are segmented to encompass their total attenuation anomaly, which is then converted to a volume based on the end-member CT numbers of the feature and the surrounding matrix. For measuring shape and orientation, we use the brightest (or darkest, for negative features such as pores) voxels up to the PVE volume. We demonstrate the method on a series of gold grains, which present additional challenges due to their irregular shapes and severe attenuation leading to scanning artifacts, scanned with two instruments at a range of resolutions and with various surrounding media. We recover volume accurately and reproducibly with few resolution effects over a factor of 27 range in grain volume and factor of five range in data resolution, successfully characterizing particles as small as 5.4 voxels in true volume. Shape metrics are affected variably by resolution effects, and are more reliable when based on caliper measurements than perimeter length or surface area. Orientations of major and minor grain axes are reproducible when their measured lengths are sufficiently different from the intermediate axis. Calibrating the PVE method requires knowledge of the end-member CT numbers for the features of interest, which are obtained empirically here; we attempt a first-principles calculation and discuss its challenges. Altogether, we find the PVE method to be an accurate, reproducible, resolution-invariant, and objective approach to measuring small features in CT data volumes, all important improvements over threshold-based methods. Threshold-based segmentation will tend to overestimate the size of features that are large relative to the data PSF, and underestimate or omit small features, to an extent that is likely to vary with sample size and data resolution.