Design of a Low-Cost Computer-Aided Diagnosis System for Knee Osteoarthritis

Knee osteoarthritis (OA) is one of the most common diseases worldwide, leading to activity limitations and physical disability in humans. Early detection and quantitative assessment of OA using the Kellgren-Lawrence (KL) grading system on plain radiographs can help slow disease progression. Currently, computer-aided methods based on deep learning can reliably assess OA severity, but they require a large amount of labeled data to train neural networks. In this work, we propose using transfer learning to achieve fully automatic detection and classification of knee joints in plain radiographs. In addition, we propose a semi-supervised learning algorithm (SSL) based on consistency regularization to leverage both unlabeled data and labeled data to improve classification performance. On public knee X-ray datasets, compared with state-of-the-art supervised learning (SL) methods, our method can achieve great effects while greatly reducing the cost of labeling data.