Osteoarthritis Severity Classification in Knee X-Rays Using Optimized Deep Learning Approaches.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Faculty of Medicine, Sivas Cumhuriyet University, Sivas, Turkey. [email protected].
- Department of Mechatronics Engineering, Faculty of Technology, Sivas Cumhuriyet University, Sivas, Turkey.
- Division of Rheumatology, Department of Physical Medicine and Rehabilitation, Faculty of Medicine, Sivas Cumhuriyet University, Sivas, Turkey.
- Department of Radiology, Faculty of Medicine, Sivas Cumhuriyet University, Sivas, Turkey.
Abstract
Osteoarthritis (OA) is a common, progressive joint disease that significantly reduces quality of life and limits mobility, especially in older adults. It is essential to accurately classify the disease in its early stages to develop effective treatments and slow its progression. This study introduces a deep learning-based system for classifying OA severity using knee joint X-ray images. EfficientNetB1, DenseNet169, and Xception architectures were employed for five-class OA classification (asymptomatic, early, mild, moderate, and severe), with hyperparameters in the fully connected layers optimized through the Gray Wolf Optimization (GWO) algorithm. By automatically selecting the most suitable parameters with GWO, the model learns more effectively and provides more accurate results in distinguishing OA levels. The dataset includes knee X-ray images from patients at University Training and Research Hospital, comprising 1000 images-200 per class. Model performance was assessed using accuracy, precision, recall, F1 score, and ROC curves. The experiments consisted of two stages: the first involved a five-class classification, and the second involved binary classification to distinguish between mild and severe OA. Correctly identifying moderate and severe stages, which are particularly serious, is vital for determining the need for surgical intervention. The best results were achieved with the DenseNet169 model: 74% in multi-class classification and 93.75% in binary classification. These findings show that the optimized models deliver high accuracy and effectiveness in OA diagnosis. This system helps specialists determine OA severity levels early, allowing for more informed treatment and surgical decisions.