A 2D-registration algorithm for the correction of motion-induced misalignments of consecutive image stacks in multi-stack high-resolution peripheral quantitative CT scans.

Abstract

Multi-stack imaging using high-resolution peripheral quantitative CT (HR-pQCT) can involve misalignments of consecutive image stacks ('stack shift') due to subject movement during scan acquisition. We developed a simple, 2D-registration algorithm for the correction of stack shifts in multi-stack HR-pQCT scans and investigated 1) the differences in standard HR-pQCT parameters and repeatability between before and after stack-shift correction; and 2) the correlation between the transformation needed for the stack-shift correction and corresponding difference in HR-pQCT parameters. The algorithm generates an artificial stack overlap of two slices, then rigidly registers the overlapping region (only in-plane translation allowed), and subsequently applies the resulting translation to the proximal stack. The algorithm was applied to data of 23 men and women with three same-day repeated scans (69 radius and 63 tibia scans, Dataset 1) and of 48 postmenopausal women with 78 radius scans taken at two time points with 12-week interval (Dataset 2). In both datasets, median differences in HR-pQCT parameters between before and after stack-shift correction were mostly significant yet small (≤0.53 %). The differences could vary considerably between subjects and ranged between -12.1 % and +35.8 % for cortical porosity, stiffness, and failure load. For the other HR-pQCT parameters, the differences ranged between ±0.8 % (Dataset 1) and between -4.5 % and +0.9 % (Dataset 2) among subjects. Spearman correlations between the magnitude of the translation and corresponding difference in HR-pQCT parameters were significant for most parameters in both datasets and strongest for stiffness and failure load (ρ = 0.687-0.947; p < 0.01). Based on Dataset 1, coefficients of variation differed between ±0.3 percentage points after stack-shift correction as compared to before. To conclude, correction of stack misalignments in two-stack HR-pQCT scans using our algorithm resulted in significant but negligible median differences in HR-pQCT parameters and precision, but differences could exceed least-significant differences and thereby be clinically relevant in individual subjects. The translation needed for the stack-shift correction correlated significantly with the difference in most HR-pQCT parameters, thereby potentially serving as objective measure for stack-shift severity. The algorithm can be applied directly after scan reconstruction, at low computational cost and without negative effects from image interpolation.