Publications

Abstract : Worldwide 1.7 billion people suffer from various musculoskeletal conditions and it leads to severe disability and long-term pain. Due to the lack of limited qualified radiologists in various parts of the world, there is a need for an automatic framework that can accurately detect abnormalities in the radiograph images. Deep learning (DL) is very popular due to its capability of extracting useful features automatically with less human intervention, and it is used for solving various research problems in a wide range of fields like biomedical, cybersecurity, autonomous vehicles, etc. The convolutional neural network (CNN) based models are especially used in many biomedical applications because CNN is capable of automatic extraction of the location-invariant features from the input images. In this chapter, we look at the effectiveness of various CNN-based pretrained models for detecting abnormalities in radiographic images and compare their performances using standard statistical measures. We will also analyze the performance of pretrained CNN architectures with respect to radiographic images on different regions of the body and discuss in detail the challenges of the data set. Standard CNN networks such as Xception, Inception v3, VGG-19, DenseNet, and MobileNet models are trained on radiograph images taken from the musculoskeletal radiographs (MURA) data set, which is given as an open challenge by Stanford machine learning (ML) group. It is the large data set of MURA that contains 40,561 images from 14,863 studies (9045 normal and 5818 abnormal studied) which represents various parts of the body such as the elbow, finger, forearm, hand, humerus, shoulder, and wrist. In this chapter, finger, wrist, and shoulder radiographs are considered for binary classification (normal, abnormal) due to the fact that data from these categories are less biased (less data imbalance) when compared to other categories. There are in total 23,241 and 1683 images given as train and valid set in this data set for the three categories considered in the present work. In the experimental analysis, the performance of the models are measured using statistical measures such as accuracy, precision, recall and F1-score.