Automatic diagnosis and localization of tuberculosis in chest radiographs with deep learning

Abstract

Tuberculosis (TB) is an infectious lung disease that is caused by Mycobacterium tuberculosis bacteria. Tuberculosis spread rapidly through airborne bodily fluids of an infected person, which has the potential to cause a serious outbreak of the disease, especially in developing countries where prevention and screening effort can be lacking. This necessitates quick and accurate diagnosis method for TB. For this reason, chest radiographs (X-ray) are usually prescribed to screen for sign of TB manifestation prior to more time-consuming methods. However, the result should be interpreted by a radiologist, which can be a problem in settings with limited resources. In this thesis, we presented the implementation of a deep learning based method for the automatic diagnosis and localization of TB in chest X-ray scans obtained from Vietnam hospitals. Two deep neural networks, based on Inception-ResNet V2 and RetinaNet, was developed and validated on four datasets from Vietnam, China, and the United States. Principle of transfer learning was used extensively to boost the performance of the networks on the small training data. Both neural networks are capable of producing diagnosis localization of tuberculosis in chest X-ray images. Our proposed method achieved an area under the receiver operating curve of 99% in the internal validation dataset, with excellent generalization in the two external validation datasets (AUROC of 99.3% and 97.5%). The localization produced by the algorithm was quantitatively and visually consistent with radiologist's manual annotations.