A new model for anomaly detection in elbow and finger X-ray images: Proposed parallel DenseNet

Image recognition is one of the essential branches of computer vision and has significant theoretical and practical importance. This study aims to enhance a deep learning model, DenseNet, by incorporating parallel structures using X-ray images from the MURA (musculoskeletal radiographs) dataset. X-ray images of the elbow and finger are analyzed using AlexNet, DenseNet, parallel DenseNet, and proposed parallel DenseNet (PPDN) deep learning models for anomaly detection, and the results are compared. For the elbow, 1534 healthy and 1630 anomaly X-ray images; for the finger, 1965 healthy and 1938 anomaly X-ray images were used to train the deep learning models. As a result of the statistical analysis, the most successful model with the test accuracy value for the elbow part was the suggested PPDN model (78.74%). The next successful model for the elbow part was AlexNet (77.05%). The most successful model for the finger part was again the PPDN model (69.97%), and the next successful model was the parallel DenseNet model for the finger part (68.94%). In anomaly detection of musculoskeletal elbow and finger X-ray images, the PPDN model is more successful than the classical DenseNet and AlexNet models in terms of test accuracy.