A Transformer-based deep learning intelligent diagnosis approach of DXA images for bone mineral density classification

Osteoporosis and osteopenia are common age-related bone diseases. Dual-energy X-ray absorptiometry (DXA) is the gold standard for measuring bone mineral density (BMD). Traditional DXA assessment relies on clinical information, commercial software, and manual correction by radiologists, making it both subjective and time-consuming. To address these issues, this study proposes BMD-TranNet, a deep learning model that integrates transformer, attention gate, and convolutional neural network (CNN) for automatic three-category classification of DXA images into normal, osteopenia, and osteoporosis. We also developed a web-based intelligent diagnosis platform using Flask for potential clinical application. A total of 2151 clinical DXA images were collected in the study, including 850 cases of osteopenia, 950 cases of osteoporosis, and 351 cases of normal bone mass. The dataset was partitioned strictly at the patient level into a training set (1,936 images) and a held-out validation set (215 images). All images were obtained from two common clinical sites (lumbar spine and left hip) with well‑defined anatomical distributions. The performance of BMD-TranNet was compared against several classic baseline models. The proposed BMD-TranNet achieved the highest overall accuracy of 0.837, outperforming all four classic baseline models: Swin Transformer(0.816), ResNet101 (0.793), DenseNet121 (0.788), and ResNet50 (0.777). The areas under the ROC curve (AUC) for osteoporosis, normal bone mass, and osteopenia for BMD-TranNet reached 0.943, 0.967 and 0.879, respectively. The developed web platform supports user login verification, local DXA image upload, real-time preview, one-click prediction through an intuitive user interface with simple operation. Experimental results demonstrate that BMD-TranNet model outperformed the existing classic baseline models. Explainable analysis shows that it can effectively capture the key feature information of DXA images from different anatomical sites by up‑weighting important channel features. The intelligent platform can serve as an auxiliary diagnosis reference for clinical practitioners. However, prospective and external validation are needed before clinical deployment.