An attention-based CNN–Transformer hybrid framework for osteoporosis and osteopenia classification in knee X-ray images

Osteoporosis is a widespread bone disease characterized by loss of bone strength, and thus, individuals, particularly elderly people and postmenopausal women, are prone to fractures. Plain radiographic images still represent a major clinical challenge in the early identification of osteoporosis, especially in the accurate distinction between osteopenia and osteoporosis due to subtle variations in trabecular and cortical bone structures. The study presents a hybrid deep learning framework that integrates the DenseNet169 model and a Vision Transformer (ViT) through an attention-based feature fusion mechanism to detect osteoporosis and osteopenia in knee X-ray images automatically. It is a dual-branch architecture that builds on the complementary capabilities of convolutional neural networks to encode fine-grained local texture detail and transformers to encode global contextual interaction within bone areas. To further enhance bone structure visibility and improve feature representation, a domain-specific preprocessing pipeline is incorporated, including bone windowing, bilateral denoising, contrast-limited adaptive histogram equalization (CLAHE), and edge-preserving sharpening. An attention-driven fusion strategy is employed, thus addressing clinically significant structural patterns in an adaptive manner, finally boosting discriminative ability. To address class imbalance, data augmentation techniques are applied, resulting in a balanced dataset across osteoporosis, osteopenia, and healthy classes. Further, the performance and effectiveness of the proposed model are evaluated by a 5-fold stratified cross-validation. Additionally, the explainable AI method, Grad-CAM, is incorporated to improve clinical interpretability, visualize salient regions, and measure the contribution of the features to the model’s decisions. Experimental results demonstrate that the developed framework achieves improved performance compared to individual backbone models and existing approaches for osteoporosis and osteopenia detection. Ablation studies further verify the effectiveness of the attention-based fusion process as compared to standalone DenseNet169, standalone ViT, and traditional feature concatenation methods. Altogether, the proposed framework has promising potential to become an interpretable computer-aided decision support system for osteoporosis screening using routine knee radiographic imaging. Routine clinical use needs to be validated on larger, multi-center clinical datasets.