Detection model for pulmonary tuberculosis and performance evaluation on histogram enhanced augmented X-rays

Abstract

Tuberculosis is one of the biggest threats that has been remaining a contagious disease since its discovery, posing a significant risk to millions of lives. Many people yield to TB because of incomplete treatments or the lack of preventive measures. An effective pulmonary TB diagnostic system has remained a big challenge. As it is a contagious disease, it mainly affects the lungs and other vital organs of the human body. We find DL as a subset of ML that runs an incurable disease diagnostic system with multi-neural architectures. In recent ages, a neural model can detect more accurately and quickly resulting in classified labels as normal and positive TB cases.    It helps medical practitioners to identify bacterial infections in the early stage. It has also enabled proper diagnosis and treatment for pulmonary tuberculosis. Through this paper, an enhanced detection model to classify TB and non-TB cases using clinical X-ray images has been proposed. The augmented histogram equalized X-rays were applied to top state-of-the-art classifiers. The evaluation matrics have been compared with and without histogram equalization and a comparative study is done to find the best CNN classifiers. The Resnet 50 and ResNet169 have shown the higest accuracy on preprocessed chest X-rays with 99.6% and 99.48% respectively.