Detection of Hand Osteoarthritis From Hand Radiographs Using Convolutional Neural Networks With Transfer Learning

Abstract

Osteoarthritis is the most common type of arthritis. Hand osteoarthritis leads to specific structural changes in the joints, such as asymmetric joint space narrowing and osteophytes (bone spurs). Conventional radiography has traditionally been the primary method of visualizing these structural changes and diagnosing osteoarthritis. We aimed to develop a computerized method that is capable of determining the structural changes seen in radiography of the hand and to assist practitioners in interpreting radiographic changes and diagnosing the disease. In this retrospective study, transfer-learning-based convolutional neural networks were trained on a randomly selected dataset containing 332 radiography images of hands from an original set of 420 and were validated with the remaining 88. Multilayer convolutional neural network models were designed based on a transfer learning method using pretrained AlexNet, GoogLeNet, and VGG-19 networks. The accuracies of the models were 93.2% for AlexNet, 94.3% for GoogLeNet, and 96.6% for VGG-19. The sensitivities of these models were 0.9167 for AlexNet, 0.9184 for GoogLeNet, and 0.9574 for VGG-19, while the specificity values were 0.9500, 0.9744, and 0.9756, respectively. The performance metrics, including accuracy, sensitivity, specificity, and precision, of our newly developed automated diagnosis methods are promising in the diagnosis of hand osteoarthritis. Our computer-aided detection systems may help physicians in interpreting hand radiography images, diagnosing osteoarthritis, and saving time.