Classification of Shoulder X-ray Images with Deep Learning Ensemble Models

Abstract

Fractures occur in the shoulder area, which has a wider range of motion than other joints in the body, for various reasons. To diagnose these fractures, data gathered from X-radiation (X-ray), magnetic resonance imaging (MRI), or computed tomography (CT) are used. This study aims to help physicians by classifying shoulder images taken from X-ray devices as fracture/non-fracture with artificial intelligence. For this purpose, the performances of 26 deep learning-based pre-trained models in the detection of shoulder fractures were evaluated on the musculoskeletal radiographs (MURA) dataset, and two ensemble learning models (EL1 and EL2) were developed. The pre-trained models used are ResNet, ResNeXt, DenseNet, VGG, Inception, MobileNet, and their spinal fully connected (Spinal FC) versions. In the EL1 and EL2 models developed using pre-trained models with the best performance, test accuracy was 0.8455, 0.8472, Cohen’s kappa was 0.6907, 0.6942 and the area that was related with fracture class under the receiver operating characteristic (ROC) curve (AUC) was 0.8862, 0.8695. As a result of 28 different classifications in total, the highest test accuracy and Cohen’s kappa values were obtained in the EL2 model, and the highest AUC value was obtained in the EL1 model.

Comparing Stacking Ensemble Techniques to Improve Musculoskeletal Fracture Image Classification

“…The main difference between our study and Uysal et al’s [ 15 ] is that we used all seven different MURA datasets. Uysal et al [ 15 ] used only the shoulder dataset.…”

“…The main difference between our study and Uysal et al’s [ 15 ] is that we used all seven different MURA datasets. Uysal et al [ 15 ] used only the shoulder dataset. Thus, their results cannot be generalized over the different types of fracture images.…”

“…Thus, their results cannot be generalized over the different types of fracture images. Additionally, compared to Uysal et al [ 15 ], we repeated each experiment 10 times to increase the significance of our findings. On the other hand, Uysal et al [ 15 ] provided no average results and only presented their findings on a single model.…”

“…Additionally, compared to Uysal et al [ 15 ], we repeated each experiment 10 times to increase the significance of our findings. On the other hand, Uysal et al [ 15 ] provided no average results and only presented their findings on a single model. They reported a kappa score exceeding our highest score for the shoulder dataset; we might speculate that this is due to the fact that they considered high-resolution images.…”

“…Uysal et al [ 15 ] studied the effect of different SOTA CNN architectures on the MURA dataset’s shoulder images. The images were scaled to 320 × 320 pixels.…”