Research Of Tuberculosis Detection In Sputum Smear Images Based On Color And Shape

TB is a severe disease damaging people’s health. China has the one of the heaviest burden of TB around the world. One thirds of Chinese are infected, over400,000,000. Statistically speaking,130,000people died of TB every year and they are averaged55.2years old. According to recently research, the number of the infected will grow to50million if not taking measures. Cell recognition plays an important role in medical image-processing, especially in the diagnosis of TB. Computer can significantly improve the efficiency of diagnosis and provide additional information. These works below are done in this thesis.1. Some observations are performed from the perspective of human vision first. Then extraction on color features are carried out, using the Vector Quantization technique to simplify the original images and color space transformation from common RGB to HSV, CIEL*a*b*, etc. After the transformation, it is noticed that red elements can be extracted from channels like Hue, L, a and Y, Cr, which accomplished the task of extraction of color feature.2. But the TB bacilli smear image is more complicated. Colors like blue background varied in the image, lightness of the image is not equal, come bacillus even cannot be recognized by human due to mixing with dark blue tissues. Moreover, when the density of TB bacillus is high enough, many will overlay on each other like parallel lines, crosses or extended lengths. To deal with these problems, Gaussian Mixture Models are introduced. GMM is a kind of unsupervised clustering method, which is easy to train and more accurate than methods like K-means. Gaussian mixture models is a mature method for clustering unknown data. To determine the parameters of GMM, we use Expectation Maximization algorithm, which uses unlabeled data for model training. The experiment shows GMM finished the initial work of TB detection, while its performance wasn’t high enough.3. To perform more comprehensive classification, Naive Bayes Classifier is built with more features like local color difference, roundness of shape and area of ROI, combining with color features. First, the Bayes Theorem is introduced to deduce the Bayes Classification Model. Considering these many features are used, the Assumption of Conditional Independence is also introduced and the Bayes Classification Model derived to the Naive Bayes Classifier. Using the Naive Bayes Classifier, the Red/Green offset from CIEL*a*b*model, hues from HSV space, shapes and areas can all be considered as features to train a Naive Bayes Classifier. The experiment result says a Naive Bayes Classifier is more accurate and robust than ordinary methods.