Computer-aided extra-pulmonary tuberculosis diagnosis using image processing and HMLP network

Abstract

Tuberculosis (TB) is a disease cause by infection with bacteria called Mycobacterium tuberculosis. The bacteria usually attack the lungs, resulting in pulmonary TB (PTB), but they can also affect other parts of the human body, referred as extra-pulmonary TB (EPTB). The disease can be cured in most people with appropriate antibiotic treatment. Unfortunately, it remains a major public health problem and the second leading cause of death worldwide among the infectious diseases. Early diagnosis of the disease is essential for effective patient treatment and preventing TB transmission in the community. Direct microscopy of sputum smear and tissue sections are the most widely used test for the PTB and EPTB diagnosis respectively. However, the diagnosis method is time-consuming, tedious, labour-intensive and has low sensitivity. Research on computer-aided detection and diagnosis of PTB that aims to overcome the problems of manual screening has attracted a lot of attentions recently. However, very little attention has been focused on the EPTB disease. Thus, this research proposes a method for automatic detection of TB bacilli in tissue sections for EPTB diagnosis. A procedure of image segmentation is proposed to segment Ziehl-Neelsen (ZN) stained tissue images into TB bacilli and background regions. The procedure combines the thresholding and clustering algorithms, and utilizes the C-Y colour model to perform the segmentation process. A local adaptive thresholding is also proposed to improve the results of segmentation. In the feature extraction stage, the research considers three types of features namely simple shape descriptors, colour information and region-based moment invariants to represent the segmented objects and used in classifying the objects. In addition, this research also proposes the skeleton-based moment invariants to improve the classification performance of TB bacilli using the conventional region-based moment invariants. Three types of skeleton-based moment invariants namely skeletonbased Hu’s, Zernike’s and affine moment invariants are formulated. In the classification stage, this research proposes to use HMLP network for detection and classification of TB bacilli. A multiclass HMLP network is designed for simplicity of training process and replacing the hierarchical HMLP (HP 2 PMLP) network. The research also rederives three activation functions; hyperbolic tangent, adaptive sigmoid and adaptive hyperbolic tangent activation function to suit for training HMLP network and replace the commonly used sigmoid functions. In addition, three types of training algorithms namely Levenberg-Marquardt (LM), modified extreme learning machine (MELM) and hybrid modified prediction error-modified extreme learning machine (MRPE-MELM) algorithms are rederived for HMLP network to improve the network performance. Finally, a software that utilized the best of above mentioned methods is proposed to perform the automated TB bacilli detection. Overall, the software provides good ability in detecting and classifying TB bacilli. The software manages to achieve promising accuracy, sensitivity and specificity of 84.25%, 85.64% and 86.98% respectively. The software also showed a high sensitivity on TB-positive images when evaluated using the image-based analysis. These findings suggest the potential use of this software in diagnosing EPTB disease, especially for patients with highly suspected EPTB infection.