Mycobacterium tuberculosis recognition with conventional microscopy

This paper presents a new method for segmentation of tuberculosis bacillus in conventional sputum smear microscopy. The method comprises three main steps. In the first step, a scalar selection are made for characteristics from the following color spaces: RGB, HSI, YCbCr and Lab. The features used for pixel classification in the segmentation step were the components and subtraction of components of these color spaces. In the second step, a feedforward neural network pixel classifier, using selected characteristics as inputs, is applied to segment pixels that belong to bacilli from the background. In third step geometric characteristics, especially the eccentricity, and a new proposed color characteristic, the color ratio, are used to noise filtering. The best sensitivity achieved in bacilli detection was 91.5%.