Knee osteoarthritis severity classification in radiographs using advanced convolutional neural networks.

BACKGROUND: Knee osteoarthritis (OA) is a degenerative, progressive joint disease with narrowing of the joint space, formation of osteophytes, and sclerosis of the subchondral regions, which results in disability and loss of mobility and quality of life. Radiographic images are critical in the assessment of OA severity to enable early diagnosis and treatment planning. Nevertheless, the traditional assessment of the Kellgren-Lawrence (KL) grading system is not objective and may have inter-observer reliability. OBJECTIVE: Current automated techniques cannot consistently predict the localization of the knee joint and have a low capability of representing features, which may impact classification. To address these constraints, this work proposes a state-of-the-art convolutional neural network (CNN)-based model to train automated knee OA severity classification based on radiographs. WORK DESIGN & METHODS: The proposed solution combines accurate localization of the knee joints and deep feature learning to improve the accuracy of classification. A YOLOv2-based model was first used to automatically locate and extract the knee joint region to guarantee accurate localization of the region of interest. A CNN consisting of AlexNet was then employed to extract and classify the features into KL grades (0-4). RESULTS: Training and evaluation were conducted using 3000 knee radiographic images, resulting in performance metrics of 85% accuracy, 84% precision, and 83% recall. CONCLUSIONS: The suggested approach enhances the reliability and consistency of OA severity measurements, indicating its potential role in facilitating clinical decision-making and early diagnosis.