Automated TB classification using ensemble of deep architectures

Tuberculosis (TB) is an infectious disease that mainly affects the lung region. Its initial screening is mostly performed using chest radiograph, which is also recommended by the World Health Organization. To help the radiologists in diagnosing this disease, different computer-aided diagnosis (CAD) systems have been developed. However, the development of these systems are still in the early phases as it is extremely challenging to automatically detect TB. This is due to extreme variations in the impact caused by TB on the CXR. In this study, a deep-learning-based TB detection system has been presented which achieves significantly high accuracy. The proposed method is an ensemble of three standard architectures namely AlexNet, GoogleNet and ResNet. The significant contribution of the study is to train these architectures from scratch and creating an ensemble suited to perform TB classification. The proposed method is trained and evaluated on a combined dataset formed using publicly available standard datasets. The ensemble attains the accuracy of 88.24% and area under the curve is equal to 0.93, which eclipses the performance of most of the existing methods.