Enhanced deep residual network for bone classification and abnormality detection

Abstract

Purpose A two‐stage deep learning framework for bone classification and abnormality detection is proposed based on X‐rays. The primary focus is on improving the speed of orthopedic disease diagnosis and helping physicians reduce the probability of false diagnoses. Methods The method is based on two stages. In the first stage, one classifier with ResNeXt50 as the backbone is used to classify bones to eliminate the effect of bone type differences on abnormality detection. In the second stage, seven anomaly detectors are trained based on each type of training data. The seven detectors tested the seven results of the first stage, respectively. Pretrained models, data augmentation, focal loss, label smoothing loss, LR‐attenuation and early stopping are used to improve performance and reduce the risk of overfitting. Results Experiments are based on the largest dataset for bone abnormality detection, MURA. In the first stage for bone classification, we got an accuracy of 96.69%, a sensitivity of 96.69%, a specificity of 99.46%, and an F1 score of 96.42%. In the second stage for abnormality detection, we got an accuracy of 84.15%, a sensitivity of 84.15%, a specificity of 87.50%, an F1 score of 84.10%, a Cohen's Kappa of 0.72, and an area under the ROC curve (AUC) score of 0.90. Conclusions Compared with other excellent convolutional neural network models, the framework's effectiveness was verified with better accuracy, sensitivity, specificity, F1 score, Cohen's Kappa score, and AUC score for bone classification and abnormality detection.