Layer-based deep net models for automated classification of pulmonary tuberculosis from chest radiographs

Abstract

Tuberculosis (TB) is a highly infectious bacterial disease. However, it can affect any body part, but is majorly a lung infection; which is potentially fatal and contagious. Like most of the serious health issues, the recovery rate of a symptomatic TB patient completely depends on the early detection and treatment. Deep learning algorithms-based computer aided diagnosis (CAD) system, can provide aid in early detection of the disease. In this regard, a method to detect infection of tuberculosis, which uses deep learning network to classify CXR images as normal or abnormal, is presented. Convolutional neural network (CNN), visual geometry group (VGG16) and high-resolution network (HRNet) models are used and their performance has been compared based on the validation loss and validation accuracy. The HRNet provides 89.7% accuracy with comparatively less loss among the proposed algorithms. The models are also deployed in android application for active clinical trials.