How Can a Deep Learning Algorithm Improve Fracture Detection on X-rays in the Emergency Room?

One of the principal reasons for attending the emergency room (ER) is peripheral traumatism. The first radiological examination in any case of suspected fracture remains the conventional X-ray. Fractures can be difficult for junior physicians to diagnose in situations of high patient flux [1,2]. The misdiagnosis of fractures directly affects patient management, and serious complications, such as malunion or arthritis, may occur if fractures are diagnosed late or remain undiagnosed. Fracture misdiagnosis is also one of the commonest causes of litigation in the domain of medicine.

The algorithm is an ensemble algorithm composed of multiple object detection models. Each object detection model is based on artificial convolutional neural network (ACNN) technology and, more specifically, is derived from the architecture of RetinaNet [14], an open-source DL algorithm. It has three components. The first is a custom backbone (based on the VGG neural network [15] but with fewer filters and batch normalization performed before each convolution) that acts as a feature extractor. This backbone consists of convolutional layers, max-pooling layers, and trainable batch normalization layers. The second component is a feature pyramid network (FPN) designed to extract the features at different resolutions, given the large variability of fracture size. The final component is two subnetworks; a classification subnet for predicting the presence or absence of a fracture, and a regression subnet for localizing the site of the fracture more precisely. The classification subnet predicts the probability of an object being present, for any class (two classes in our case), at each spatial position, for each anchor. The classification subnet is applied to each pyramid level, but the parameters of this subnet are shared across all pyramid levels. The classification subnet is a fully convolutional network. The regression subnet predicts the offset from each anchor box to a nearby ground-truth object (if such an object exists). We also applied this regression subnet to each pyramid level (with shared weights). This subnet is also a fully convolutional network. It is similar to the classification subnet except that, rather than predicting two (i.e., the number of classes) values, it has four values per anchor (Figure 1).

In total, the emergency physicians included 125 patients in this study. The traumatism considered concerned the hip in sixteen patients, the hand in twenty-eight, the shoulder in twenty-eight, the foot in twenty one, the knee in seven, the wrist in twenty one, and the elbow in four.

This first study on a population attending the ER for peripheral traumatism shows that a DL algorithm can be used, with a high level of accuracy, including a high negative predictive value in particular. A DL algorithm with no training could be used at a new center without the need for data from this site, for the diagnosis of fractures in a population of patients consulting for traumatisms of any peripheral joint.