Kwantitatieve beeldvorming van bot met klinische fotonentellende computertomografie

Quantitative imaging involves the extraction of quantifiable features from medical images, which can be used when diagnosing or monitoring disease, injury or chronic conditions. Computed tomography (CT) is the most appropriate imaging modality for three-dimensional visualization of bone; quantitative analyses from CT can be used in the disease management of bone degenerative diseases such as osteoporosis, osteoarthritis and rheumatoid arthritis. High-resolution peripheral quantitative CT (HR-pQCT) is the gold standard for in vivo quantitative imaging of bone; however, it can only scan small sections of distal radius and tibia, for which a long scanning time is needed (10.2 mm in 2 minutes), which limits its clinical applicability. In contrast, clinical CTs can provide full body scans in roughly one minute, yet, at a resolution typically not sufficient for high measurement precision. Photon-counting CT (PCCT) is a novel clinical CT modality, with increased resolution and image quality compared to conventional CT. This provides new perspectives for site-specific, high-resolution quantitative analyses, directly applicable in a clinical context. This thesis aimed to assess the feasibility of quantitative imaging of bone with clinical PCCT. More specifically, the goals were 1) to quantify the accuracy of bone mineral density (BMD) quantification with PCCT, relative to HR-pQCT; 2) to quantify the accuracy and short-term precision of bone microarchitecture quantification with PCCT, relative to HR-pQCT and micro-CT (µCT); 3) to quantify the accuracy of voxel-based finite element (FE) models from PCCT; 4) to evaluate different methods for three-dimensional analysis of joint space morphology, and assess their clinical utility. For this, we acquired a multimodal CT dataset of eight cadaveric forearms that were scanned twice with PCCT, once with HR-pQCT, and after which the extracted carpal bones were scanned with µCT. Additionally, we scanned ten samples from bovine vertebrae with PCCT, conventional clinical CT, and µCT. Paired measurements between PCCT and HR-pQCT demonstrated high accuracy when measuring average BMD within a volume of interest (R²=0.972, mean bias 7.46 mgHA/cm³). The repeated PCCT scans showed a high reproducibility of the voxel grayscale value; from this we calculated that the least detectable amount of bone formation or resorption that can be measured at the microscopic level constitutes 250 mgHA/cm³. This is clinically relevant when compared to fully mineralized bone (1200 mgHA/cm³), and aligns closely with the least detectable remodeling with HR-pQCT (225 mgHA/cm³). When comparing the microarchitecture of carpal bones from PCCT and HR-pQCT, and relative to µCT, both HR-pQCT and PCCT showed measurement error, which often also varied linearly with the measured value. However, using an adaptive threshold for PCCT segmentation resulted in an accuracy and agreement that approached the results of HR-pQCT. Additionally, when calculating microarchitecture parameters from adaptive threshold segmentation in PCCT, short-term precision was higher than when measured from the standard global threshold segmentation. Image-based quantifications of bone strength and stiffness at the distal radius from PCCT and HR-pQCT were highly similar (R² > 0.977, bias 0.046 kN). Equally high agreement was found for the proportional load in the cortical and trabecular bone compartments. For vertebral samples, there was no bias between FE-based quantifications of apparent linear-elastic parameters from µCT and PCCT; in contrast, FE models from conventional CT had significant bias, and a lower resulting anisotropy. When using non-linear density-scaled material models to quantify the vertebral samples' ultimate failure load, PCCT correlated more strongly to experimental results than conventional CT. When comparing two frequently used methods for joint space morphology on various articulations in the hand and wrist, there were differences in results, which originated from fundamental differences in interpretation of joint space width. From our limited data, joint space morphology analyses from grayscale deconvolution methods were more sensitive to reflect progressive radiographic damage than those from joint space segmentation methods. Finally, we also evaluated the accuracy of various algorithms for trabecular microstructure segmentation, and proposed a new conceptual method based on a Gaussian mixture model (GMM). For our data, adaptive thresholding of PCCT led to the highest total accuracy; yet, the GMM provides a parametrization that can favor a higher sensitivity (i.e. segmenting more fine details) or a higher specificity (i.e. a more precise delineation of the segmented structures). To conclude, we presented high accuracy of various quantitative analyses with PCCT, relative to HR-pQCT, µCT and experimental measurements. Future work should reevaluate whether or not standard HR-pQCT methodologies are directly transferable to new imaging modalities, and evaluate the added value of these analyses in medical workflows.