Deep learning analysis of anomaly detection from musculoskeletal radiography images

Musculoskeletal disorders are anomalies of the bones and muscles that affect the majority of the world's population. Radiographic studies are the most common technique for detecting these anomalies as part of medical diagnosis. Deep learning, a sub-unit of artificial intelligence today, is being applied in many areas of complex and critical importance in the world. In this study, deep learning methods, whose classification studies continue to increase, are used to detect musculoskeletal anomalies. The aim of this study is to develop a deep learning model, DenseNet, using parallel layers with X-ray images from the MURA (Musculoskeletal Radiographs) dataset. In the study, the preprocessing stages were first applied to X-ray images of the hand, finger, wrist, forearm, elbow, humerus and shoulder in the MURA dataset. X-ray images were rotated, projected on axes and scaled to increase the amount of input image data. X-ray images with the same resolution and bit depth were trained with a convolutional neural network (CNN) using MATLAB software and classified as healthy and anomaly. In order to develop the classical DenseNet method in the study, X-ray images were trained by adding parallel blocks in the MATLAB program and using the Parallel DenseNet and Proposed Parallel DenseNet (ÖPDN) methods. Additionally, classification was made with AlexNet and ResNet, a deep learning method, to compare the results. The same parameter settings were used in all ESA models. X-ray images were trained in ESA with an 80% training, 15% test, and 5% validation ratio using a random data selection method. In the classification made with these four ESA architectures; accuracy, precision, sensitivity, specificity, F1-score, Cohen's kappa statistic and k-fold cross-validation performance measurements were used. As a result of the study, the values of the test accuracy of the ÖPDN model for the hand part 66.16%, finger part 69.97%, wrist part 73.86%, forearm part 74.07%, elbow part 78.74%, humerus part 78.65% and shoulder part 68.42% were more successful than the other AlexNet, ResNet, classical DenseNet and Parallel DenseNet models. The results obtained show that the ÖPDN model can be used successfully for early and accurate diagnosis in classification, in analyzes made with MURA X-ray images.