Year: 2026
Type: dissertation
Abstract: This dissertation explores two distinct aspects of osteoarthritis (OA) research: first, the evaluation of deep learning models for detecting structural abnormalities in knee MRIs, and second, the identification of knee pain phenotypes using statistical modeling techniques. In the first study, deep learning models were developed to detect structural abnormalities, particularly bone marrow lesions (BMLs), in knee MRIs using imbalanced data. A total of 2996 sagittal intermediate-weighted fat-suppressed knee MRIs from 2467 participants in the Osteoarthritis Initiative (OAI) study were utilized. The analysis revealed that commonly used evaluation metrics, such as the receiver operating characteristic (ROC) curve, were insufficient when applied to imbalanced data. Specifically, in regions with highly imbalanced class ratios, the ROC curve was not informative enough, while the precision-recall curve (PR) provided a better measure of model performance. The results underscore the importanceof reporting multiple evaluation metrics, including ROC, PR, and Matthews correlation coefficient (MCC), when assessing deep learning models trained on imbalanced data. In the second study, we examined knee pain trajectories by combining four datasets, increasing the sample size from 2560 to 3066 knees. Group-Based Multi-Trajectory Modeling (GBMTM) and Latent Class Mixed Models (LCMM) were used to identify knee pain progression. The union of the datasets allowed for a more robust analysis, which identified three distinct knee pain trajectories: minimally progressive, slowly progressive, and rapidly progressive. Logistic regression analyses revealed that bone marrow lesions (BMLs), higher BMI were associated with a higher likelihood of being in the rapidly progressive group. Notably, gender is insignificant in both analysis, which validated previous findings. These results emphasize the importance of MRI features such as BMLs, cartilage surface damage in predicting knee pain progression, while also highlighting the role of demographic factors in shaping pain trajectories. In sum, these studies provide valuable insights into both the statistical modeling of knee pain trajectories and the evaluation of deep learning methods for OA research. The use of appropriate evaluation metrics for deep learning models ensures more accurate assessments of model performance, particularly in the context of imbalanced data. Furthermore, the integration of imaging biomarkers in longitudinal studies of knee pain trajectories can enhance early diagnosis and personalized treatment. These findings suggest the need for further research that combines both statistical modeling and advanced machine learning techniques to better understand and predict the progression of knee OA.
Author: Liu, Shen
Institution: University of Arizona
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7162161315