Key points are not available for this paper at this time.
BACKGROUND: Although the automatic diagnosis of fractures using artificial intelligence (AI) has recently been reported to be more accurate than those by orthopedics specialists, big data with at least 1000 images or more are required for deep learning of the convolutional neural network (CNN) to improve diagnostic accuracy. The aim of this study was to develop an AI system capable of diagnosing distal radius fractures with high accuracy even when learning with relatively small data by learning to use bi-planar X-rays images. METHODS: VGG16, a learned image recognition model, was used as the CNN. It was modified into a network with two output layers to identify the fractures in plain X-ray images. We augmented 369 plain X-ray anteroposterior images and 360 lateral images of distal radius fractures, as well as 129 anteroposterior images and 125 lateral images of normal wrists to conduct training and diagnostic tests. Similarly, diagnostic tests for fractures of the styloid process of the ulna were conducted using 189 plain X-ray anteroposterior images of fractures and 302 images of the normal styloid process. The distal radius fracture is determined by entering an anteroposterior image of the wrist for testing into the trained AI. If it identifies a fracture, it is diagnosed as the same. However, if the anteroposterior image is determined as normal, the lateral image of the same patient is entered. If a fracture is identified, the final diagnosis is fracture; if the lateral image is identified as normal, the final diagnosis is normal. RESULTS: The diagnostic accuracy of distal radius fractures and fractures of the styloid process of the ulna were 98. 0 ± 1. 6% and 91. 1 ± 2. 5%, respectively. The areas under the receiver operating characteristic curve were 0. 991 n = 540; 95% confidence interval (CI), 0. 984-0. 999 and 0. 956 (n = 450; 95% CI 0. 938-0. 973). CONCLUSIONS: Our method resulted in a good diagnostic rate, even when using a relatively small amount of data.
Oka et al. (Thu,) studied this question.