Abstract
Osteoporosis is a prevalent bone disease characterized by reduced bone density and an elevated risk of fractures, especially in older adults and postmenopausal women. The clinical consequences of osteoporotic fractures extend beyond pain and disability, contributing substantially to morbidity, mortality, and healthcare costs. Early intervention and accurate detection is therefore essential to improve patient outcomes. This paper introduces an advanced deep-learning methodology to enhance the accuracy and efficiency of osteoporosis detection through knee X-ray analysis. The proposed approach integrates features from two pre-trained models, DenseNet169 and Vision Transformer (ViT), with a custom-designed Attention Model (AM) to capture detailed spatial and channel-specific information from the input images. These fused features are then fed into a fully connected neural network to classify the images as osteoporotic or normal. The results indicate significant improvements in classification accuracy, achieving a high accuracy rate on previously unseen test data. The proposed model achieves superior performance over existing methods and other recent models for osteoporosis detection, with an accuracy of 0.8611, specificity of 0.9474, and precision of 0.9286. Our approach effectively combines convolutional and transformer-based representations, enabling extraction of both local and global features for comprehensive bone characterization. These findings highlight the model's potential to support early diagnosis, timely intervention, and improved patient care in osteoporosis management.