Abstract:
Knee osteoarthritis (KOA) is a common degenerative joint disorder that impacts millions worldwide, particularly older adults and is a significant cause of chronic pain and mobility impairments. Spotting KOA early isn't easy because the early signs on X-rays can be subtle, and even experts using the standard KL grading system might see things differently. In this research, I explored whether deep learning could help identify early knee osteoarthritis from standard X-ray images. The researcher set up a system to distinguish between definitely normal knees (KL Grade 0) and definite early-stage osteoarthritis (KL Grade 2). KL Grade 1 (doubtful/minimal changes) was excluded from model training to eliminate diagnostic ambiguity. The study focused on this clinically meaningful binary distinction for early-stage detection. For this work, 3,000 knee X-rays from the Osteoarthritis Initiative were used, making sure to balance healthy and early arthritis cases. The approach was to test three different deep learning models side-by-side: one I built from scratch (a custom CNN), plus MobileNetV2 and InceptionV3. After training them all, InceptionV3 consistently gave the best results, so that became my final model choice
The InceptionV3 approach came out on top, hitting 97% accuracy in tests. What's really encouraging is that it showed strong sensitivity and ROC performance, meaning it's good at catching true cases without too many false alarms. Data augmentation was applied exclusively to the training set to enhance generalization without biasing validation or test evaluation. To improve model transparency and clinical trust, Gradient-weighted Class Activation Mapping (Grad-CAM) was used. The resulting visualizations highlighted anatomically relevant regions such as the medial joint compartment and joint space narrowing in KL Grade 2 cases, which matches up with established radiographic biomarkers of KOA.
The results show that accurate and clinically significant early diagnosis of knee osteoarthritis from X-ray images can be supported by deep learning in conjunction with interpretability approaches. This method could help radiologists make better decisions and enhance early intervention techniques.
