High-Resolution Quantitative Cone-Beam Computed Tomography: Systems, Modeling, and Analysis for Improved Musculoskeletal Imaging

This dissertation applies accurate models of imaging physics, new high-resolution imaging hardware, and novel image analysis techniques to benefit quantitative applications of x-ray CT in in vivo assessment of bone health. We pursue three Aims: 1. Characterization of macroscopic joint space morphology, 2. Estimation of bone mineral density (BMD), and 3. Visualization of bone microstructure. This work contributes to the development of extremity cone-beam CT (CBCT), a compact system for musculoskeletal (MSK) imaging. Joint space morphology is characterized by a model which draws an analogy between the bones of a joint and the plates of a capacitor. Virtual electric field lines connecting the two surfaces of the joint are computed as a surrogate measure of joint space width, creating a rich, non-degenerate, adaptive map of the joint space. We showed that by using such maps, a classifier can outperform radiologist measurements at identifying osteoarthritic patients in a set of CBCT scans. Quantitative BMD accuracy is achieved by combining a polyenergetic model-based iterative reconstruction (MBIR) method with fast Monte Carlo (MC) scatter estimation. On a benchtop system emulating extremity CBCT, we validated BMD accuracy and reproducibility via a series of phantom studies involving inserts of known mineral concentrations and a cadaver specimen. High-resolution imaging is achieved using a complementary metal-oxide semiconductor (CMOS)-based x-ray detector featuring small pixel size and low readout noise. A cascaded systems model was used to performed task-based optimization to determine optimal detector scintillator thickness in nominal extremity CBCT imaging conditions. We validated the performance of a prototype scanner incorporating our optimization result. Strong correlation was found between bone microstructure metrics obtained from the prototype scanner and µCT gold standard for trabecular bone samples from a cadaver ulna. Additionally, we devised a multiresolution reconstruction scheme allowing fast MBIR to be applied to large, high-resolution projection data. To model the full scanned volume in the reconstruction forward model, regions outside a finely sampled region-of-interest (ROI) are downsampled, reducing runtime and cutting memory requirements while maintaining image quality in the ROI