Advancing Osteoporosis Diagnosis through State-of-the-Art CNNs and Vision Transformers with Ensemble Strategies

Abstract

Osteoporosis is a common bone disorder marked by reduced mineral density and microarchitectural deterioration, increasing fracture risk. Early, accurate detection is vital for clinical intervention and personalized care. This study applies deep learning to binary classification of osteoporosis (normal vs. osteoporotic) using medical imaging. A curated dataset of bone-related images was used to train and evaluate advanced models and ensemble strategies. Evaluated architectures include EfficientNetB2, InceptionV3, InceptionResNetV2, ResNet50V2, Xception, Vision Transformer (ViT_B32), and Faster R-CNN. Accuracy served as the main metric. ResNet50V2 outperformed all with 97.83% accuracy, ahead of EfficientNetB2 and ViT_B32 (95.65%), InceptionV3 and Xception (95.22%), InceptionResNetV2 (93.91%), and Faster R-CNN (76.96%). Ensembles—average, weighted, and hard voting—further improved accuracy to 96.96% and 96.09%. The results validate the benefit of ensemble learning in boosting model robustness. ResNet50V2 stands out as the top single model, and ensemble techniques show strong promise for reliable, automated osteoporosis detection. These findings support deploying deep learning in clinical radiology for early diagnosis and decision support.