Osteotrans-Net: An interpretable transformer framework for knee osteoarthritis severity grading in X-ray scans

Knee osteoarthritis is a degenerative joint disease that commonly affects older adults, causing pain and stiffness. Kellgren–Lawrence (KL) grading is a manual process, skill-based, and prone to human error. This paper proposes Osteotrans-Net, a hybrid architecture that integrates Vision Transformers with CNNs for accurate detection and severity classification of knee osteoarthritis (KOA). The input X-ray images are initially preprocessed and then passed through a 3D CNN, which captures both spatial and temporal features using multiple convolutional layers. The Dual Focus Attention (DFA) module improves feature representation by highlighting essential areas. The reduced features are fused with the original data using 2D convolution, batch normalization, and the ReLU activation function in the fusion block. The enriched attributes are processed through a pre-trained ViT, succeeded by a classifier that designates the final label—either healthy or osteoarthritic knee. We utilize the “Knee Osteoarthritis Dataset” and divide it into three distinct subsets: Dataset-I, Dataset-II, and Dataset-III. Osteotrans-Net model has exceptional performance in experimental assessments across three distinct datasets, attaining classification accuracies of 89.57%, 94.38%, and 89.88% respectively. We compare our model with three pre-trained deep learning models-ResNet18, EfficientNetB0, and ViT. Comparative analysis shows our strategy. The model’s attention and interpretability for KOA are assessed using Grad-CAM. Our study highlights how integrating attention mechanisms with feature extractors significantly enhances model performance, as revealed by ablation experiments showing improved feature learning and accuracy. Statistical analysis confirms that our strategy leads to improved KOA diagnosis. Osteotrans-Net is available on https://github.com/000jd/Osteotrans-Net-Interpretable-Transformer-for-Knee-OA-Grading-in-X-ray to promote reproducibility and further study. • Proposed Osteotrans-Net integrates 3D CNN, DFA & ViT for KOA severity grading. • DFA improves spatial and channel-wise focus, boosting feature representation. • Achieves 89.6–94.4% accuracy on 3 datasets, surpassing ResNet18, EffNet-B0 & ViT. • Grad-CAM shows critical joint regions, enhancing clinical interpretability.