An Expert-Assisted Chest Radiograph Classification Tool To Differentiate Between Active Tuberculosis and Healed Tuberculosis Sequelae

Abstract

Tuberculosis (TB) heals with a scarring response in the lung and chest radiographic abnormalities termed as tuberculosis sequelae are often observed in patients with healed pulmonary TB. Patients with healed TB can also have frequent chest infections, breathlessness and fever, which are also seen in active pulmonary TB. This leads to frequent mismanagement of these patients where they are treated as active TB cases. In this paper, we present a method to differentiate between active and healed pulmonary TB patients using their chest radiograph images. Our method requires the clinician to acquire the image of the chest radiograph using a hand-held cellular smartphone, and then mark points along the boundary of the lungs in the image. We then use these point markings in an iterative Otsu-based thresholding method, parameterized optimally using a discrete simulation optimization technique, to segment the lungs from the rest of the image. Clinically relevant morphological features such as the lung areas, lung base curvatures and tracheal shift, and textural features such as average intensity within the lung, are extracted from the segmented image. We then train multiple standard machine learning methods on the extracted features, and convolutional neural networks on the segmented images themselves to classify the radiograph images. We achieve area under the receiver operating characteristic curve scores of up to 0.83. We also present results from preliminary experiments conducted to estimate the extent of inter-doctor variability in marking points on the lung boundaries.