Automated Diagnosis of Vertebral Fractures Using Radiographs and Machine Learning

Objective: People often experience spinal fractures. The most common of these are thoracolumbar compression fractures and burst fractures. Burst fractures are usually unstable fractures, often accompanied by neurological symptoms, and thus require prompt and correct diagnosis, usually using computed tomography (CT) or magnetic resonance imaging (MRI). However, X-ray images are the cheapest and most convenient tool for predicting fracture morphological patterns. Therefore, we built a machine learning model architecture to detect and differentiate compression fractures from burst fractures using X-ray images and used CT or MRI to verify the diagnostic outcome.Methods: We used YOLO and ResUNet models to accurately segment vertebral bodies from X-ray images with 390 patients. Subsequently, we extracted features such as anterior, middle, and posterior height; height ratios; and the height ratios in relation to fractures and adjacent vertebral bodies from the segmented images. The model analyzed these features using a random forest approach to determine whether a vertebral body is normal, has a compression fracture or has a burst fracture.Results: The precision for identifying normal bodies, compression fractures, and burst fractures was 99%, 74%, and 94%, respectively. The segmentation and fracture detection results outperformed those of related studies involving X-ray images.Conclusion: We believe that this study can assist in accurate clinical diagnosis, identification, and the differentiation of spine fractures; it may help emergency room physicians in clinical decision-making, thereby improving the quality of medical care.