Deep Learning for Bone Health-A Transformer Based Framework for Osteoporosis Screening

Abstract

Osteoporosis is a progressive bone disease that leads to a higher risk of fractures, and early diagnosis continues to be a significant problem because current diagnostic tools are not comprehensive. In this work, we tested deep learning models, such as VGG16, VGG19, ResNet, DenseNet, UNet, Vision Transformer (ViT), and Swin Transformer, on 853 X-ray images to detect osteoporosis automatically. To overcome the issue of class imbalance and enhance the robustness of the model, data preprocessing was carried out through standardization, normalization and augmentation. It was experimentally demonstrated that VGG19 had the best classification accuracy of 97%, and UNet performed better in a segmentation task with a Dice score of 0.936. In comparison with the available literature, our method shows a competitive level of performance and proves the possibility of CNN-based and transformer models in medical image processing. The present work helps to build trustworthy, inexpensive, and affordable AI-aided diagnostic systems to monitor osteoporosis and establishes the future trends towards larger datasets, multimodal systems, and real-time clinical use.