Characterizing Phenotypes of Musculoskeletal Degeneration Using Medical Imaging and Deep Learning

Musculoskeletal disease is the leading cause of disability worldwide, with the 2019 Global Burden of Disease study reporting global disease prevalence of approximately 1.714 billion1. X- ray and magnetic resonance imaging (MRI) are routinely used for clinical diagnosis and monitoring of musculoskeletal disease, however, due to an increasing volume of acquired images and limited time, image assessments are mainly qualitative. This thesis aims to elevate the role of imaging in the assessment of musculoskeletal disease by developing fully automatic image analysis tools to improve image analysis sensitivity, speed, and/or precision. We target the two conditions with the highest prevalence and healthcare expenditure in the United States: knee osteoarthritis (OA) and back pain. We use deep learning to develop fully automatic tools for image analysis and demonstrate their utility in the assessment and analysis of research and clinical datasets. I will be presenting four main projects:(1) A deep learning segmentation method for quantitative analysis of knee cartilage from structural MRI to conduct longitudinal analysis on cartilage thickness over 8 years (2) A point cloud algorithm for feature learning from structural and compositional knee MRI to assess the importance of shape and composition features in predicting OA onset (3) A registration pipeline for voxel-based analysis of MR imaging of the lumbar spine to examine local associations between T1ρ, T2, and patient reported outcomes (4) A curve extraction algorithm for analysis of global spine shape from x-ray imaging to build a shape model that examines 3D spine shape variations in the UCSF patient populatio