Representation learning in knee osteoarthritis

The primary aim of this thesis was to develop novel deep learning-based methods to advance the KOA field in two crucial aspects: data efficiency and clinical relevance. First, for data efficiency, a semi-supervised method was developed, which requires a significantly small number of annotated samples and leverages unannotated ones to perform KOA diagnosis. Second, for clinical relevance, deep learning was utilized to tackle KOA prognosis prediction, which is more clinically relevant than the KOA diagnosis task as it allows forecasting the severity of the disease in the future. Finally, the thesis concludes by raising a question regarding the confidentiality risk of participants in representation learning, particularly within the context of KOA.