Automatic opportunistic osteoporosis screening using chest X-ray images via deep neural networks

Comprehensive Summary

This study provides a new deep learning (DL) architecture that uses chest X-rays to automatically screen for osteoporosis. A dataset of 1870 images was compiled, with 300 being used for testing, 1184 for training, and 296 for model validation. Three pretrained deep learning models (Inception v3, VGG16, and ResNet50) were selected and adjusted to screen for osteoporosis. Additionally, a basic convolutional neural network (CNN) was used as a baseline framework. Model performance was evaluated based on Area under the Curve (AUC), accuracy, sensitivity, and specificity. The mean accuracies for the CNN, Inception v3, VGG16, and ResNet50 were 0.804, 0.859, 0.861, and 0.904, respectively. The ResNet50 achieved the highest overall performance (AUC= 0.957, Sensitivity= 0.918, Specificity= 0.887). Despite the strong results shown, the model also had 15 false positive and 13 false negative cases. Overall, results show that with further refinement and fine-tuning, DL models may be an efficient way to identify osteoporosis from X-rays.

Outcomes and Implications

The current “gold standard” for osteoporosis identification (DXA) faces many limitations such as high costs, radiation exposure, and equipment penetration. Using DL models to read chest X-ray images provides a more efficient and streamlined alternative. This method has the ability to save time for clinicians, resources for health systems, and can protect patients from repetitive radiation exposure. By potentially identifying osteoporosis earlier, deep learning models may also lead to better patient outcomes and more specified treatment options. Further research and expansion will lead to the improvement of model performance, allowing this method to be used in more clinical environments.