Diagnosis and Analysis of Tuberculosis Disease Using Simple Neural Network and Deep Learning Approach for Chest X-Ray Images

Abstract

AbstractArtificial Intelligence (AI) based diagnosis of Tuberculosis (TB) disease has experienced large developments with the application of Machine Learning (ML) and Deep Learning (DL) methods to classify the disease as well as to detect it. TB is a contagious infection that usually presents on lungs and is identified through the initial level symptoms by conducting tests using Chest Radiographs (CXRs) and microscopic images. Powerful Diagnosis of Tuberculosis depends on rigorous analysis of radiological patterns realized in CXR. However, due to high number of patients burden and lack of resources in underdeveloped country is high chance of human error in analyzing the CXRs and hence, the diagnosis of TB becomes difficult. Our aim is to develop a computer aided diagnosis (CAD) system for TB disease classification, which can help in early diagnosis of the disease. Nowadays, deep learning based automatic feature extractors are used, when large dataset is concerned for accurate classification instead of using hand crafted features. Our work deals with above mentioned methods to have a justified explanation when working with small- and large-scale data learning problems. This work proposes two approaches for classification of TB disease using X-Ray dataset. In first approach, we have utilized handcrafted features in simple neural network with Support Vector Machine (SVM) to classify the disease as TB images and normal images. We have designed ANN system for 13 input neurons, 10 hidden neurons and 2 output neurons to train features, which are fed as input to SVM for classification purpose. Experiments are conducted on MATLAB 2014b. ANN-SVM based classification gives accuracy of 94.6% when all features are fed to it. This method is thereby awarding increase in efficiency and reduced diagnosis time. In second, we implemented Deep learning technique, which is capable of training high level features from dataset compared to handcrafted feature method to classify the TB disease. In which we have developed binary classification using Deep Convolutional Neural Network (DCNN). Google collab notebooks are used to model DCNN with GPU based Keras library and Tensor flow as back end. Experiments are conducted on Tuberculosis Chest X-ray dataset obtained from Kaggle community and showed output classification accuracy of 99.24%.