Accuracy and precision of segmentation and quantification of wrist bone microarchitecture using photon-counting computed tomography ex vivo

The quantification of bone microarchitecture provides insight into bone health and the effects of disease or treatment, and is therefore highly relevant clinical information. Nonetheless, in vivo quantification of bone microarchitecture is mostly limited to high-resolution peripheral quantitative CT (HR-pQCT). This is a small field of view CT modality of which the gantry size only allows scanning of distal radius and tibia. Photon-counting CT (PCCT) is a novel clinical full-body CT with improved image resolution and quality compared to other clinical CT modalities, yet data on its capabilities in quantifying bone microarchitecture are limited. The aim of this study was to quantify the accuracy of two methods for trabecular bone segmentation on PCCT images as compared to the segmentations on micro-CT (µCT) and to use these segmentations to quantify the accuracy and agreement of trabecular bone morphometry measurements as compared to µCT, as well as the short-term precision. This study analysed multimodal CT data, obtained from eight cadaveric forearms; the data includes two repeated PCCT scans, as well as a single HR-pQCT scan from the forearm, and µCT scans of all individual carpal bones. For each carpal bone, trabecular volumes of interest (VOI) were delineated on the µCT images, and the µCT reference segmentations and VOIs were resampled onto the PCCT and HR-pQCT images. HR-pQCT images were segmented with a global threshold of 320 mgHA/cm3; PCCT images were segmented with either an identical global threshold or with an adaptive thresholding algorithm. Trabecular bone-volume fraction (Tb.BV/TV), trabecular thickness (Tb.Th), trabecular number (Tb.N) and trabecular separation (Tb.Sp) were quantified for all segmented VOIs. Accuracy and agreement were calculated relative to µCT as the gold standard, short-term precision was calculated from the repeated PCCT scan. For PCCT, adaptive threshold segmentation had significantly increased sensitivity compared to global threshold segmentation, along with a lower variance in its sensitivity and specificity. Concerning the microarchitecture quantification, for global threshold segmentation of PCCT images, correlations with µCT were significant, except for Tb.Sp. Correlation coefficients of Tb.BV/TV and Tb.N were not significantly different from those between HR-pQCT and µCT. Adaptive threshold segmentation led to higher correlation coefficients between PCCT and µCT of Tb.Th, Tb.N and Tb.Sp, although correlations of Tb.N remained poor for both PCCT and HR-pQCT. Moreover, adaptive thresholding led to a constant bias of Tb.BV/TV, Tb.Th and Tb.Sp, unlike the bias of HR-pQCT which was proportionally increasing with the size of the measurement. Finally, adaptive threshold segmentation led to a higher short-term precision than global threshold segmentation, with a root-mean-squared coefficient of variation below 0.65 % for all parameters. We conclude that adaptive threshold segmentation is well-suited for the segmentation of PCCT images. Despite measurement error, our results indicate that these segmentations can be used for bone microarchitecture analyses of carpal bones with agreement and short-term precision comparable to HR-pQCT.