AI-driven model for knee cartilage degeneration using SAM, Swin, Grad-CAM, and CapsNet

Knee cartilage degeneration is a major contributor to mobility disability, and affects millions of individuals worldwide. For effective intervention, early accurate detection of the condition is essential. Conventional approaches often lack sufficient sensitivity, whereas recent artificial intelligence (AI) techniques have demonstrated improved diagnostic accuracy. The study aims to develop an AI-based model for accurate classification of knee cartilage degeneration stages, integrating segmentation, feature extraction, and classification techniques to address existing limitations. The proposed model employs the Segment Anything Model (SAM) for segmentation, the Swin Transformer for feature extraction, and a Capsule Network (CapsNet) for classification.The Dice Similarity Coefficient (DSC) and Intersection Over Union (IoU) were used to evaluate the segmentation's performance. For classification, accuracy, precision, recall, and F1-score were used. Gradient-weighted class activation mapping (Grad-CAM) was employed to support interpretability. Segmentation achieved a DSC of 0.90 and IoU of 0.83, enabling accurate extraction of cartilage regions. Grad-CAM correlation values ranged from 0.85 to 0.89, indicating consistent localization of relevant cartilage regions. Classification accuracy reached 96.5% for healthy, 94.8% for mild, 95.2% for moderate, and 93.7% for severe cases, with corresponding F1-scores ranging from 93.0% (severe) to 96.0% (healthy). Overall, the proposed model achieved an accuracy of 97%, with a precision of 96% and a recall of 95%, demonstrating robust performance across all evaluation metrics. Future work includes expanding dataset diversity, exploring multimodal approaches, and implementing real-time clinical applications to enhance diagnostic accuracy and personalized treatment planning.