Impact of Optimization Algorithms in Detection of Tuberculosis using Transfer Learning

Abstract

The World Health Organization has identified
tuberculosis (TB) as a major public health threat. Delayed
diagnosis due to factors like lack of experts and highly
accurate diagnostic methods that are relatively expensive
contributes to the increasing and the prevailing burden of TB
in the developing countries. This may can be addressed with
the breakthroughs in Artificial Intelligence (AI) which paved
the way for Deep learning to specialize in computer vision
fields. The Convolutional Neural Networks (CNN), a
promising algorithm, has shown potential in medical image
recognition applications as effective models for extracting
relevant features from images since it does not demand
objective-specific manual feature engineering and promotes
end-to-end training from extracting features to classification.
The limitation of publicly accessible datasets, on the other
hand, is a challenge which can be resolved utilizing the
Transfer Learning technique. This study uses transfer
learning on chest X-rays (CXRs) to detect tuberculosis, from
which features are extracted using VGG-19 architecture
pretrained on ImageNet and then fed into a classifier for
prediction. We prepared and analyzed the performance of the
network through accuracy (ACC), area under the ROC curve
(AUC) and confusion matrix on Shenzhen dataset. The
influence of three different optimizers on the result obtained
was assessed and compared.