Performance Analysis of Different Convolutional Neural Network (CNN) Models with Optimizers in Detecting Tuberculosis (TB) from Various Chest X-ray Images

Abstract

Tuberculosis (TB) is one of the top 10 infectious disease-related deaths. This paper uses Convolutional Neural Networks (CNN) to investigate the accuracy and performance of three pre-trained models with different optimizers and loss functions to diagnose tuberculosis based on the patient's chest X-ray scans. The odds of treating and curing tuberculosis (TB) are better if the disease is diagnosed early in a patient. Early detection of tuberculosis could lead to a decreased overall mortality rate. The best and quickest way to identify tuberculosis is to look at the patient's chest X-Ray image (CXR). A qualified professional Radiologist is required to make an accurate diagnosis. But do not have qualified doctor or radiologist everywhere. On the other hand, it is quite difficult for a doctor or radiologist to diagnose from any x-ray images with open eyes. 914 normal chest x-ray images and 892 TB infected images were used from different sources to train and evaluate these images to detect the exact x-ray of Tuberculosis infected people. Different famous pre-trained models like VGG16, InceptionV3 and Xception etc. were applied. Approximately 80% of the data was used for training and the remaining 20% was used for validation. From all of these datasets, randomly 190 images from normal and 180 images from TB chest x-ray images have been taken. Those randomized 370 (190 for TB and 180 for normal) images were used to evaluate the data finally. Performance of different algorithm like VGG16, InceptionV3 and Xception by applying different optimizers (Adam, Adadelta, Adagrad, Adamax, RMSprop, Nadam, SGD), different loss functions (Binary Cross Entropy, Hinge, Squared Hinge), varying input image size and also varying batch size were also been recorded. Note that, huge variations of performance for different combinations of algorithm, optimizer, loss function, input image size, batch size have been observed. Confusion matrix, precision, recall, f1-score value have also been recorded to understand and justify how accurately the model is predicting the disease from different angles.