Segmentation And Classification Of Optical Microscopic Mycobacterium Tuberculosis Images

Tuberculosis(TB)is a chronic infectious disease caused by Mycobacterium tuberculosis(MTB).Early diagnosis and treatment play an important role in the control and cure of the TB.In the background of computer-aided diagnosis of the TB,this dissertation focuses on the segmentation and recognition methods of optical microscopic MTB images.The primary objective is to solve several key problems existing in the segmentation and recognition of the MTB images.The main contents and contributions are summarized as follows:To solve the problem that the MTB objects have low contrast in the images,in Chapter 2 we improved the traditional histogram equalization,and proposed a color image enhancement algorithm based on the weighted thresholded histogram equalization(WTHE).The proposed algorithm was successfully applied to the MTB images,and produced good results with contrast enhancement,artifacts suppression,and brightness preserving.The experimental results showed the effectiveness of the proposed algorithm.To solve the background color diversities and the incomplete object segmentation of the MTB images,a segmentation algorithm based on the globalized probability of boundary(gPb)watershed was proposed in Chapter 3.With this algorithm,the MTB objects were segmented in the perspective of region segmentation,which reduced the influence of color variations of the object pixels.The multi-thresholds segmentation was proposed to deal with the images in different background colors.In order to avoid over-segmentation and improve the efficiency of the algorithm,a segmentation algorithm based on markerbased watershed was proposed.To precisely locate the markers of the objects and reduce the interference with the markers from the contaminations around the objects,we proposed a markers-obtaining method based on an adaptive scale Gaussian filter.The experimental results demonstrated that the proposed algorithm improved the accuracy of segmentation.Meanwhile,the algorithm obtained high robustness for the images in different background colors.To deal with the broken objects existing in the segmentation results,a skeleton-based connection algorithm was proposed in Chapter 4.In this algorithm,the skeleton information was used to connect the ends of the broken objects.In the meantime,the close objects were not connected by mistake,which improved the effectiveness of the algorithm.To deal with the touching MTB objects,a skeleton-based segmentation algorithm was proposed.The single object was extracted from the touching objects.Three basic methods were presented to analyze the one-branch-point skeletons.A hierarchy tree analysis method was proposed for the multi-branch-point skeletons.Several one-branch-point skeletons can be separated from the multi-branch-point skeletons based on the hierarchy tree.To deal with close branch points,a fusion method was proposed.The experimental results indicated that the proposed methods solved the problems of broken objects connection and touching objects segmentation.This work has provided foundation for the following object recognition.To solve the problem in the feature selection of the MTB object recognition,a recognition algorithm based on active shape model(ASM)was proposed in Chapter 5.Unlike the commonly-used geometrical features and descriptors,the deformation regulation of the object shape was obtained by establishing the shape model of the MTB objects.The recognition rules were determined by the deformation regulation.First,a point distribution model of the object shape was proposed to automatically label the landmarks on the objects.Second,the algorithm of generalized procrustes analysis(GPA)was used to align the training set,and the weight of the landmarks was added to reduce the aligning errors.The principal component analysis(PCA)was utilized to obtain the regulation of the shape variation.Finally,a recognition method based on shape parameter and width consistency constraint was proposed.The experimental results demonstrated the effectiveness of the proposed algorithm.