Knee Osteoarthritis Classification From X-Ray Scans Using Deep Learning and Grad-CAM

Abstract

Knee osteoarthritis (KOA) is a common type of arthritis that causes chronic joint pain and impaired mobility. Early and accurate classification is essential for timely intervention and effective treatment planning. This study presents a web-based decision support system for classifying KOA severity using the Kellgren-Lawrence (KL) grading system. The Knee Osteoarthritis with Severity Grading dataset from Kaggle was used, consisting of 8,260 anteroposterior X-ray images after excluding the ‘auto-test’ subset. Preprocessing involved CLAHE, normalization, and targeted data augmentation for underrepresented classes, with class weights applied to mitigate class imbalance. The system employs pre-trained convolutional neural networks (CNNs), specifically ResNet50, VGG16, and DenseNet121, to analyze knee X-ray images. A total of 243 model configurations were evaluated by varying optimizers, learning rates, batch sizes, and epoch sizes. Among the tested models, the optimal configuration, VGG16 with SGD optimizer, learning rate 0.0001, batch size 8, and 30 epochs, achieved 43.60% accuracy, 25.57% precision, 30.01% recall, and an F1-score of 27.08%. Per-class evaluation showed low to moderate performance for KL grades 0 to 2, while grades 3 and 4 yielded near-zero scores. Grad-CAM heatmaps were integrated for interpretability but often failed to highlight clinically relevant regions, reflecting weak feature localization. Despite these limitations, the system offers an accessible tool for KOA assessment and provides a foundation for future improvements through enhanced class balancing, feature extraction, ordinal regression, and model fine-tuning.