Densenet201+ with multi-scale attention and deep feature engineering for automated Kellgren–Lawrence grading of knee osteoarthritis

Accurate and early diagnosis of knee osteoarthritis (KOA) is critical for effective treatment and patient management. In this study, I propose an enhanced deep learning-based diagnostic framework centered on our custom-designed DenseNet201+ architecture, specifically optimized for automated Kellgren–Lawrence (KL) grading from radiographic images. DenseNet201+ introduces architectural innovations including spatial attention mechanisms and multi-scale pooling strategies, enabling comprehensive extraction of both global anatomical patterns and localized joint features. To further enhance diagnostic performance, the methodology implements a deep feature engineering pipeline that transforms the extracted 512-dimensional feature vectors into enriched representations through higher-order statistical analysis, entropy computation, and activation modeling. These engineered features are then classified using multiple classical machine learning algorithms. Among them, a support vector machine (SVM) with a radial basis function (RBF) kernel achieved the best performance. Evaluated on a dataset of 13,254 knee X-rays with balanced class distributions across KL grades, the proposed method attained an accuracy of 94.68%, an area under the curve (AUC) of 99.40%, and a perfect AUC of 100% for severe KOA cases (Grade 4). The proposed framework demonstrates robust generalization with a cross-validation stability of 98.71% ± 0.32, and excellent inter-rater agreement (Cohen’s kappa = 0.933, intraclass correlation coefficient (ICC) = 0.951). Explainability was addressed via Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations, highlighting diagnostically relevant regions. In summary, DenseNet201+ combined with deep feature engineering and classical classification establishes a state-of-the-art, interpretable, and computationally efficient solution for automated KOA grading. Its performance exceeds that of existing models and aligns with clinical requirements, offering a scalable tool for real-time deployment in radiology workflows and remote healthcare systems.