طبقه بندي شدت بيماري آرتروز زانو با استفاده از تصاوير راديوگرافي

Knee osteoarthritis (OA) is a leading cause of physical disability worldwide, making early diagnosis critical for effective disease management. However, accurate diagnosis in the early stages—based on radiographic images an‎d the Kellgren–Lawrence (KL) grading system—remains challenging even for specialists due to its subjective nature an‎d subtle tissue variations. This study proposes an intelligent deep learning-based system for the automated detection an‎d severity grading of knee OA. To this end, various Convolutional Neural Network (CNN) architectures, including VGG, EfficientNet, an‎d DenseNet, were implemented an‎d eva‎luated. To further enhance accuracy an‎d robustness, a stacking ensemble architecture was developed, in which the outputs of base models are fused using a Support Vector Machine (SVM) as the meta-classifier. Furthermore, Grad-CAM visualization an‎d Principal Component Analysis (PCA) were employed to ensure model interpretability an‎d investigate the root causes of misclassification. The proposed ensemble achieved an accuracy of 70.05% in five-class classification, outperforming individual base models an‎d several state-of-the-art (SOTA) methods. Visual analysis confirmed that the model correctly localizes the joint space using a cropping strategy; however, severe feature overlap between KL Grades 0 (normal) an‎d 1 (doubtful) remains the primary bottleneck. Finally, by adopting a two-stage screening strategy that distinguishes healthy individuals from definite OA patients (effectively mitigating the impact of ambiguous Grade 1 cases), the system’s accuracy improved to 87.2%, highlighting its strong potential as a reliable computer-aided diagnosis (CAD) tool in clinical settings.

نادر كريمي

شادرخ سماوي

فرزانه شايق بروجني , مينا اميري