“Automatic Quantification of Radiographic Knee Osteoarthritis Severity and Associated Diagnostic Features using Deep Convolutional Neural Networks” A. Joseph Antony Due to the increasing prevalence of knee Osteoarthritis (OA), a debilitating kneejoint degradation, and total joint arthoplasty as a serious consequence, there is a need for effective clinical and scientific tools to assess knee OA in its early stages. This thesis investigates the use of machine learning algorithms and deep learning architectures, in particular convolutional neural networks (CNN), to quantify the severity and clinical radiographic features of knee OA. The goal is to offer novel and effective solutions to automatically assess the severity of knee OA achieving on par with human accuracy. Instead of conventional hand-crafted features, it is proposed in this thesis that automatically learning features in a supervised manner can be more effective for fine-grained knee OA image classification. The main contributions of this thesis are as follows. First, the use of off-the-shelf CNNs are investigated for classifying knee OA images through transfer learning by fine-tuning the CNNs. Second, CNNs are trained from scratch to quantify the knee OA severity optimising a weighted ratio of two loss functions: categorical cross entropy and mean-squared error. Third, CNNs are jointly trained to quantify the clinical features of knee OA: joint space narrowing (JSN) and osteophytes along with the KL grades. This improves the overall quantification of knee OA severity producing simultaneous predictions of KL grades, JSN and osteophytes. Two public datasets are used to evaluate the approaches, the OAI and the MOST, with extremely promising results that outperform existing approaches. In summary, this thesis primarily contributes to the field of automated methods for localisation and quantification of radiographic knee OA.