Revealing Trabecular BoneArchitecture in 3D with Deep Learningbased Super Resolution

Sammanfattning

The three-dimensional microarchitecture of trabecular bone determines itsmechanical strength, flexibility, and metabolic function, making it central toskeletal health. Changes in this structure can weaken the bone, increase fracture risk, and serve as an indicator of disease. However, the spatial resolution ofstandard clinical computed tomography (CT) scanners is insufficient to resolvethe features of bone microstructure, while higher-resolution modalities are unsuitable for routine clinical use due to limitations such as restricted imagingsites, small fields of view, and, most critically, high radiation dose. The aim of this thesis was to enable characterization of trabecular bone microstructure beyond the spatial resolution achievable with clinically availableCT scanners, using diffusion-based super-resolution. Performance was evaluated using standard trabecular bone metrics, including bone volume fraction(BV/TV), trabecular thickness (Tb.Th), trabecular separation (Tb.Sp), and trabecular number (Tb.N). Initial tests used synthetically generated low-resolutiondata to simulate different downsampling factors, up to those matching the gapto clinical resolution. The evaluation was then extended to real quantitative CT(QCT) images, a modality used in routine clinical practice. Whereas previousstudies have focused on smaller resolution gaps and relied solely on syntheticdata, this work also assessed clinically relevant resolution differences and included evaluation on real clinical scans. The results show that the model can accurately recover trabecular structureacross resolution differences up to the gap between high-resolution scans andclinical CT. Models trained exclusively on synthetic data did not generalize toreal QCT images, showing that such data do not sufficiently capture clinicalimage characteristics. When QCT data were included in training, the modelachieved accurate BV/TV and Tb.Th values. This approach holds strong potential for integrating microstructural analysis into clinical workflows, enablingearlier and more precise fracture risk assessment, personalized treatment planning, and long-term monitoring of bone health.