Attention-based Graph Convolutional Neural Network for Classification of Musculoskeletal Radiograph Images

Abstract

Musculoskeletal Disorders (MSDs) are the abnormalities related to bones and muscles, affecting majority of the world population. Radiographic studies are the most common technique for the detection of these abnormalities as part of the medical diagnoses. An attention-based graph convolutional neural network (AGCNN) is implemented, in this thesis work, for the classification of such abnormalities in musculoskeletal radiograph images. The AGCNN network model is firstly implemented on the standard benchmark MURA dataset, consisting of 40,561 upper extremity radiograph images, for the binary classification of radiograph images into normal and abnormal. The performance of the network model is compared with that of the DenseNet169 baseline model. The network model showed improved performance results than the baseline model. The network model is then implemented on Xtremity dataset, consisting of 15,701 extremity radiograph images, for the multi-class classification of radiograph images into five different classes. The network model, that is implemented, is an ensembled network of soft attention-based Inception-ResNet-v2 network and graph convolutional network (GCN). Soft Attention map is used to localize the abnormality regions in the radiograph images for qualitative evaluation of the network. The network model achieved an accuracy of 0.884, average recall of 0.874, average F1 score of 0.876, and average AUC score of 0.976. The network model achieved above average results in the classification task. Furthermore, the performance results of the classification task by the ensembled AGCNN network are compared with that of different state-of-the-art pre-trained CNN architectures.