Development of an artificial intelligence-based decision support algorithm in the radiological diagnosis of bone fractures

INTRODUCTION AND AIM: Support systems are needed to ensure timely and accurate diagnosis of fracture cases, which are common and increasing in frequency. This research aims to create an artificial intelligence-based decision support algorithm to reduce the time taken for physicians to evaluate direct radiographs of patients who apply to the emergency department with suspected fractures and to reduce the risk of missing a fracture. METHOD: 1040 fracture and 250 non-fracture radiographs were included in the study. Of these, 832 fracture radiographs and 200 non-fracture radiographs were used as training data, while 208 fracture radiographs and 50 non-fracture radiographs were used as test data. 10 different images were obtained for each image by applying image processing techniques such as adding noise, blurring, breaking, flipping and rotating to the raw radiograph images. RESULTS: 258 graphs in the test set were predicted with the created model. The model predicts 82% of patients with fractures as fractures (Sensitivity). It predicts 88% of healthy patients as healthy (Specificity). 97% of patients that the model predicts to have a fracture actually have a fracture (Precision). 54% of the patients predicted by the model to be healthy are actually healthy (Negative predictive value). The model predicts 12% of healthy patients to have a fracture (False positive rate). The model predicts 18% of fracture patients as intact (False negative rate). The frequency of the model making correct predictions among all patients is 83%. The F1 score of the model is 0.89 CONCLUSION: According to the model data created, it will enable the physician to make faster decisions, especially in cases where he thinks there is a fracture, with 97% accuracy. The use of similar models in emergency departments will minimize the frequency of incorrect and delayed diagnosis. Keywords: Plain radiography, fractures, machine learning, artificial intelligence