Classification of Osteoporosis and Detection of Compression Fractures by Implementing Real-Time Object Detection System Upon Medical Radiographs

Aim: Artificial intelligence and machine learning have been increasingly employed in medical image diagnosis, but face the challenge of database acquisition. Hence, this study applies the You Only Look Once (YOLO) technique, a real-time object detection system, to construct the inference model, including the image preprocessing and labeling, model training, and the prediction of unknown data to detect objects. Methods: We implement the YOLOv4 technique to classify osteoporosis and detect compression fractures. Trabecular characteristics of osteoporosis are extracted from caput femoris in X-ray images, and compression fractures are observed in lateral spine images. All the datasets are derived from the clinical practice of the collaborated teaching hospital. Results: We construct the YOLOv4 model to classify osteoporosis and detect compression fractures with prediction accuracy of 78.1% and 68.3%, respectively. X-ray could be a screening tool to predict osteoporosis and select patients for DXA, especially in settings where the DXA facility is unavailable. Conclusion: We find it promising to apply the developed approach to medical diagnosis with an accuracy of near 80%, and this deep learning model could preliminarily help to screen possible positives from abundant radiographs.