An Ensemble of Deep Transfer Learning Frameworks for Automatic Tuberculosis Detection in Chest X-Ray Images

Tuberculosis (TB) is a persistent pulmonary disease caused by bacterial infection and is among the ten most prevalent causes of mortality. It is crucial to diagnose TB early since the disease can be fatal if left untreated. Due to technological developments and the availability of medical datasets. An automated system for analysing and classifying chest X-rays (CXR) into TB and non-TB could potentially serve as a dependable substitute for the subjective evaluation performed by medical practitioners. A considerable segment of a CXR image is devoid of diagnostically significant data and is therefore potentially a source of confounding for deep learning (DL) models. CXR image segmentation is performed using the U-Net model; the results of the segmentation are subsequently inputted into the DL models. A classification task was executed on the Montgomery and Shenzhen datasets utilising an ensemble of multiple convolutional neural network (CNN) models, with an assessment of their performance. The proposed ensemble algorithm achieved a higher accuracy of 98.94%. The experimental results show that ensemble learning on segmented lung CXR images produces better results than unsegmented lung CXR images.