Classification of Fracture and Normal Shoulder Bone X-Ray Images Using Ensemble and Transfer Learning With Deep Learning Models Based on Convolutional Neural Networks

Various reasons cause shoulder fractures to occur, an area with wider and more varied range of movement than other joints in body. Firstly, images in digital imaging and communications in medicine (DICOM) format are generated for shoulder via Xradiation (Xray), magnetic resonance imaging (MRI) or computed tomography (CT) devices to diagnose and treat such fractures. Shoulder bone Xray images were classified and compared via deep learning models based on convolutional neural network (CNN) using transfer learning and ensemble learning in this study to help physicians diagnose and apply required treatment for shoulder fractures. There are a total of 8379, 4211 normal (negative, nonfracture) and 4168 abnormal (positive, fracture) 3 channel shoulder bone Xray images with png format for train data set, and a total of 563, 285 normal and 278 abnormal 3 channel shoulder bone Xray images with png format for validation and test data in classification conducted using all shoulder images in musculoskeletal radiographs (MURA) dataset, one of the largest public radiographic image datasets. CNN based built deep learning models herein are; ResNet, ResNeXt, DenseNet, VGG, Inception and MobileNet. Moreover, a classification was also performed by Spinal fully connected (Spinal FC) adaptations of all models. Transfer learning was applied for all these classification procedures. Two different ensemble learning (EL) models were established based on performance of classification results obtained herein. The highest Cohens Kappa score of 0.6942 and highest classification test accuracy of 84.72% were achieved in EL2 model, and the highest AUC score of 0.8862 in EL1.