Deep Learning-Based Osteoporosis Classification from Knee Radiographs Using ResNet50, Xception, and InceptionV3 Architectures

Abstract

Bones serve a crucial function in providing strength to the body and protecting vital human organs. There are at least two bone problems, namely osteopenia and osteoporosis. However, there is a problem with early detection due to high costs and lack of facilities. Therefore, this research is necessary to distinguish between those at risk and those who are not. This research presents a retrospective analysis of a publicly available knee radiograph dataset. This study uses the ResNet50, Xception, and InceptionV3 methods with data acquisition, data preprocessing, and augmentation techniques, then model building, training, validating, testing with (N = 195 images) using accuracy, precision, recall, and f1-score measurements using macro-averaging across classes, visualizing them into a confusion matrix and ROC Curve, and obtaining the AUC score results. The results of this study indicate that Xception is the method that achieved the highest accuracy of 88%, macro-precision of 88%, macro-recall of 87%, and macro-F1-score of 87%. However, in the medical world, the AUC score is considered a parameter to consider. The research results indicate that, based on the AUC scores, the three models exhibit comparable performance. The AUC score obtained for the normal class is 0.99, while for the other classes, it ranges from 0.96 to 0.97.