Deep Learning Approach Based on a Patch Residual for Pediatric Supracondylar Subtle Fracture Detection

Abstract

Abstract Background Labeled fracture radiographs are usually difficult to acquire, especially for the small sample sizes of the supracondylar fractures for children. Convolutional neural network-based methods, which rely heavily on a large amount of labeled data, cannot yield satisfactory performance. Compared to the fracture data, normal radiographs without the need for annotation are much easier to capture and include many images. Methods In this study, we propose a subtle supracondylar fracture detection framework, called the multiscale patch residual (MPR), which can learn the bone characteristics from normal data to repair fracture areas and forms residual features with the original image for fracture location. Our proposed MPR framework is a repair-based method that can learn the distribution of normal data by removing occlusions. A multiscale generation adversarial model is proposed for learning the bone consistency features from normal data. For the repaired image to be more similar to the real data, edges and textures are added as auxiliary information. Furthermore, weighted-binary cross-entropy (W-BCE) is used in our detection model to further enhance the fracture detection accuracy by adjusting the difference between the fracture area and the nonfracture area and forcing the model to learn the feature changes before and after repair. Additional experiments are conducted on the cross time independent test set, and a comparative experiment was conducted with an emergency specialist and a pediatric radiologist. The experimental results confirm the effectiveness of our approach. Results The final accuracy of independent test set was 93.5%, the sensitivity was 89%, the specificity was 98%, and the F1 value was 0.932. The accuracy of emergency physicians was 81%, the sensitivity was 62%, the specificity was 42%, and the F1 value was 0.62. The accuracy of children's radiologists was 93%, the sensitivity was 92%, the specificity was 94.2%, and the F1 value was 0.929. Conclusions Our method has achieved a good diagnostic rate, far exceeding that of emergency doctors and reaching the same level as pediatric radiologists.