Abstract:

Tuberculosis is the second infectious disease causing death after COVID-19. Diagnosing it is an easy and cheap via chest radiographs. However, some countries lack medical personnel and equipment for tuberculosis detection on chest radiographs. Computer-aided diagnosis and computer-aided detection systems utilizing deep learning can be employed to identify tuberculosis on medical images. Although there are some studies, they are insufficient for unbiased systems because these systems require the datasets having different features. The aim of this study is to evaluate the performance of pretrained networks for a classification application on chest X-ray images by utilizing the dataset from the Hospital in Turkey and Montgomery Count Dataset. The predictive models were implemented with the pre-trained DCNNs such as ResNet-50, Xception, and GoogLeNet. An Xception model provides the best performance.

Keywords: