Intelligent identification of osteoporosis on hip X-rays using vision transformer

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Bone Health Assessment
• Radiomics and Quantitative Imaging

Background:

• Osteoporosis is a major public health concern, leading to increased fracture risk.
• Early and accurate diagnosis of osteoporosis is crucial for effective management and prevention of fractures.
• Current diagnostic methods may have limitations in accessibility and cost-effectiveness, particularly in primary care settings.

Purpose of the Study:

• To develop and evaluate a deep learning model using the Vision Transformer (ViT) architecture for classifying hip X-ray images.
• To assess the model's capability in distinguishing between normal bone mass, osteopenia, and osteoporosis.
• To explore the potential of ViT for early osteoporosis screening and auxiliary diagnosis.

Main Methods:

• Retrospective collection of 3016 hip X-ray images from a hospital and community clinics.
• Preprocessing and extraction of proximal femur regions of interest (ROI).
• Fine-tuning a pretrained ViT model for three-class classification and comparison with ResNet50 and InceptionV3, validated on internal and external datasets.

Main Results:

• The ViT model achieved 97.0% accuracy on the internal validation set, with high AUCs for all classes (99.4%-99.9%).
• External validation demonstrated 89.4% accuracy, with AUCs ranging from 91.6% to 98.4%.
• The ViT model outperformed ResNet50 and InceptionV3 in diagnostic performance and stability.

Conclusions:

• The ViT-based deep learning model demonstrates excellent performance and generalizability for classifying bone mineral density from hip X-rays.
• This approach offers a cost-effective and efficient method for osteoporosis screening using routine X-ray images.
• The model holds significant potential for clinical implementation, particularly in primary care settings for early detection.