Image segmentation and object classification for automatic detection of tuberculosis in sputum smears

• Main Author: Khutlang, Rethabile
• Other Authors: Douglas, Tania S
• Format: Dissertation
• Language: English
• Published: University of Cape Town 2014
• Subjects: Biomedical Engineering
• Online Access: http://hdl.handle.net/11427/8967

Includes bibliographical references (leaves 95-101). === An automated microscope is being developed in the MRC/UCT Medical Imaging Research Unit at the University of Cape Town in an effort to ease the workload of laboratory technicians screening sputum smears for tuberculosis (TB), in order to improve screening in countries with a heavy burden of TB. As a step in the development of such a microscope, the project described here was concerned with the extraction and identification of TB bacilli in digital images of sputum smears obtained with a microscope. The investigations were carried out on Ziehl-Neelsen (ZN) stained sputum smears. Different image segmentation methods were compared and object classification was implemented using various two-class classifiers, for images obtained using a microscope with 100x objective lens magnification. The bacillus identification route established for the 100x images, was applied to images obtained using a microscope with 20x objective lens magnification. In addition, one-class classification was applied the 100x images. A combination of pixel classifiers performed best in image segmentation to extract objects of interest. For 100x images, the product of the Bayes’, quadratic and logistic linear classifiers resulted in a percentage of correctly classified bacillus pixels of 89.38%; 39.52% of pixels were incorrectly classified. The segmentation method did not miss any bacillus objects with their length in the focal plane of an image. The biggest source of error for the segmentation method was staining inconsistencies. The pixel segmentation method performed poorly on images with 20x magnification. Geometric change invariant features were extracted to describe segmented objects; Fourier coefficients, moment invariant features and colour features were used. All two-class object classifiers had balanced performance for 100x images, with sensitivity and specificity above 95% for the detection of an individual bacillus after Fisher mapping of the feature set. Object classification on images with 20x magnification performed similarly. One-class object classification using the mixture of Gaussians classifier, without Fisher mapping of features, produced sensitivity and specificity above 90% when applied to 100x images.