Improving Diagnosis of Severe Paediatric Lymphobronchial Tuberculosis Using Segmentation and Deep Learning on Chest X-Rays

Background: Diagnosing intrathoracic tuberculosis (TB) in children under five years of age remains challenging. Clinical symptoms often overlap with those of other common childhood illnesses, and microbiological confirmation is frequently difficult to obtain in this age group. Despite advances in diagnostic technologies, few studies focus specifically on children younger than five years, the population in whom diagnosis is most complex. Intrathoracic lymphadenopathy on chest X-ray (CXR) is the hallmark radiological features of TB in this age group, but identification by trained physicians currently demonstrates only moderate diagnostic performance, with reported sensitivities ranging from 67% to 74% and specificities between 39% and 59%. Objective: We aimed to develop and evaluate an automated deep learning–based approach for identifying severe pulmonary lymphobronchial tuberculosis (LBTB) in children under five years of age using anterior–posterior (AP) chest X-ray images, with confirmed airway compression serving as the primary radiographic reference. Methods: A total of 402 AP chest X-ray images were included in this study. Half of the images were obtained from children diagnosed with intrathoracic TB and with confirmed airway obstruction on bronchoscopy and/or chest computed tomography (CT). The remaining half of images were normal AP chest X-rays from children without TB disease. Three convolutional neural network (CNN) architectures—VGG16, ResNet50, and InceptionV3—were evaluated and compared for their ability to classify pulmonary LBTB based on radiographic features. Results: A fine-tuned ResNet50 model achieved a statistically significant mean test sensitivity of 94.43% and mean test specificity of 94.79% in identifying TB-positive cases. The model classified disease presence by detecting radiographic manifestations of LBTB affecting the trachea and main bronchi on CXR images.