UCSF

Knee osteoarthritis is a degenerative musculoskeletal disorder marked by gradual cartilage breakdown, and involving all tissues of the joint. It is one of the most common worldwide causes of chronic disability in older populations, with the prevalence expected to increase. There is currently no available treatment to reverse the degenerative damage characteristic in osteoarthritis, with the only option available for end stages of the disease being a partial or total knee replacement. Furthermore, the clinical standard for osteoarthritis diagnosis is a radiographic score which reflects advanced pathological stages, often with irreversible damage. The lack of therapies has generated a need for osteoarthritis imaging biomarkers capable of detecting and monitoring the progression of the disease. This dissertation aims to bridge this gap by defining a novel spherical encoding representation for known quantitative imaging biomarkers for osteoarthritis. In this work, we leverage the superior cartilage sensitivity of MRI, a large retrospective labeled imaging dataset, and the superior feature-learning ability of convolutional neural networks to define novel OA imaging biomarkers based on spherical maps. Large-scale quantitative analysis using convolutional neural networks uncovered new associations between bone shape, cartilage thickness, and cartilage T2 relaxation time values and OA symptoms.