UCSF

With the aging of the general population, musculoskeletal (MSK) diseases have moved to the forefront of healthcare concerns and are the leading causes of disability globally. Noninvasive imaging is routinely utilized in the clinic to diagnose and monitor onset and progression of MSK conditions. However, due to the qualitative nature of imaging assessments and increasing labor costs of evaluating advanced imaging modalities, there is a crucial need for automatic quantitative approaches. In this dissertation, we explore the development of computer vision techniques for extracting morphological features associated with low back pain and knee osteoarthritis, two of the most prevalent and debilitating MSK conditions.

We begin by addressing the costs of image annotation via automation with deep learning. More specifically, we developed convolutional neural networks for two purposes: (1) to semantically segment various tissues, allowing for geometric tissue characterization, and (2) to detect and localize lesions and abnormalities. Then, leveraging these models for feature extraction, we harmonized tissue geometries in 3D Euclidean space using atlas-based registration to identify tissue shapes predisposed to disease onset. These techniques were applied to both large-scale and small, limited datasets, demonstrating the utility of computer vision techniques for morphological evaluation in a data-driven, exploratory manner.