A Pruned Dense Convolution Neural Network for Detecting Tuberculosis using X-Ray Images

Artificial Intelligence: Deep Learning

Abstract

Tuberculosis (TB) is now the leading infectious disease killer in the Kenya (USAID) and the fifth leading cause of death according to Ministry of Health,Kenya report. Despite being preventable and treatable, TB continues to devastate countless susceptible populations households and communities in Kenya and across Africa. The common test for TB using skin test and blood test are not sufficient in detecting the type of the TB. The World Health Organization (WHO) recommends broader use of screening by chest X-Rays (chest radiography). However, there is a relative lack of radiology interpretation expertise in many TB prevalent locations in Kenya, which may impair screening effectiveness and work-up efforts. An effective automated and cost-effective method could aid screening evaluation efforts in Kenya and facilitate earlier detection of disease. Therefore, this study aims to develop a TB model for detecting the TB. The model employed pruning technique to minimize the size of the inference model. The model that is pruned is a dense Convolution Neural Network (DenseNet) and the resultant model is a miniature model of size 29MB against the original DenseNet of size 84MB. Our model (PDenseNet) size is approximately 3X smaller than the original DenseNet with the similar performance accuracy.