Radiographic Variability-Adaptive Structural Harmonization (RV-ASH): An Adaptive Deep Learning Framework for Robust Knee Osteoarthritis Detection and Clinical Validation

Abstract

Knee Osteoarthritis (KOA) is one of the most common degenerative musculoskeletal disorders that significantly affects joint mobility, physical function, and overall quality of life worldwide. Accurate and early-stage diagnosis of KOA from radiographic images remains challenging due to radiographic variability, heterogeneous imaging conditions, disease progression patterns, and inter-observer inconsistencies. To address these challenges, this study proposes a novel hybrid deep learning framework named Radiographic Variability-Adaptive Structural Harmonization (RV-ASH) integrated with DenseNet121 and an attention mechanism for robust KOA detection and severity classification. The RV-ASH framework reduces inter-dataset radiographic variability through adaptive structural harmonization, while DenseNet121 extracts hierarchical deep features and the attention mechanism emphasizes clinically relevant anatomical regions such as joint space narrowing and osteophyte boundaries. The proposed model was trained using the Osteoarthritis Initiative (OAI) and externally validated on the Mendeley Knee OA Dataset. Experimental evaluation achieved 98.43% accuracy, 98.15% precision, 97.92% recall, 98.03% F1-score, and 99.01% AUC, demonstrating superior classification performance, robustness, and clinical applicability for automated KOA diagnosis and severity assessment.