A novel tuberculosis diagnosis approach using feed-forward neural networks and binary pattern of phase congruency

Abstract

Tuberculosis is a severe and contagious lung disease that kills more than one million people annually, representing a real public health problem worldwide. Meanwhile, researchers and industry have highlighted the potential of analyzing and classifying chest radiographs (CXR) with artificial intelligence (AI). However, although high-performance industry solutions to aid in diagnosing have become a reality, vulnerable populations and underdeveloped countries have struggled to access them. In this sense, accessible and low-cost alternatives are essential to reach the needs of those populations that rely on public actions and services. Therefore, we propose a method to assist in the screening, diagnosing, and classifying of tuberculosis cases with a low-cost and high-efficiency computational approach. Our method uses a novel feature extraction approach based on binary patterns of phase congruence (BPPC) to describe CXR images that are represented in a feature vector of 1062 values. The extraction strategy maintains and recovers descriptive textural information of the TB, captured in 6 different directions of the CXR image, being invariant in the illumination and contrast of the CXR. The vector is then used to build a feed-forward neural network (FFNN) model capable of identifying and classifying TB cases. Our model experimented with three classification scenarios with CXR images from the TBX11K public dataset, and our results achieved exceptional performance, with greater than 99% accuracy, outperforming models from related work.

Keywords

• Artificial neural network
• Tuberculosis detection
• BPPC
• FFNN
• Chest radiograph