Knee Osteoarthritis (OA) is a prevalent and debilitating condition, impacting millions globally. While no cure currently exists, accurate and timely diagnosis is crucial for effective management and improved patient outcomes. Magnetic Resonance Imaging (MRI) provides valuable insights into soft tissue structures, aiding in OA assessment. However, current diagnostic approaches using MRI often rely on single-view analysis, potentially overlooking crucial information present in other views.

The work of my thesis aims to develop and evaluate novel deep learning models for robust and comprehensive knee OA diagnosis from multi-view MRI data. Addressing the limitations of existing methods, we propose and investigate various deep learning architectures, including multi-view, multi-label, multi-instance, multi-modal, and multi-task learning approaches. These models are designed to effectively leverage the rich information contained within multiple MRI views, enabling more accurate and holistic assessments. To further enhance performance, we designed efficient preprocessing techniques to optimize the quality of input data for both 2D and 3D deep learning models. Our proposed models, encompassing both 2D and 3D CNN architectures, were trained and evaluated on the Osteoarthritis Initiative (OAI) database, where each MRI includes multiple views (axial, coronal, and sagittal) as well as different sequences (DESS, IW-TSE, COR-MPR, AXE-MPR). In addition to diagnostic classification, we trained segmentation models to delineate specific anatomical structures relevant to OA assessment. Results demonstrate significant improvements in diagnostic accuracy, achieving accuracies ranging from 93.20 to 96 %.