Automatic Recognition On True Color Leukocyte Microscope-images

The clinical examination on the hematopoietic cells from the blood and bone marrow smears with Wright is essential for the homeopathy diagnosis and classification. The hematopoietic disease is a kind of life-threatening disease. As environmental pollution, chemical waste and the effects of physical factors, the number of people suffering from such diseases has been increasing in recent years. So it is of great significance for the early diagnosis and classification. The conventional approaches are blood cell examination (i.e.number, morphology, etc)from peripheral blood and bone smears by optical microscope. Generally in blood analysis, different categories of blood cells are distinguished by their size and color. This purely visual observation has the problem of poor reproducibility between the observers and themselves. Therefore, it is necessary to develop an automatic cell classifying and diagnostic system to complete a quantitative analysis and computer-aided detection of blood cell.With the technological development of the digital image processing, the pattern recognition and the artificial intelligence, computer-aided diagnosis has been shown to be a promising approach as a pathologists' assistant for improving diagnostic accuracy. The automation of leukocyte microscope image analysis and recognition is a representative subject in the biomedical image processing.Based on the 24-bit true color leukocyte image, this article conducts an in-depth research and study on how to use the technology of computer image processing and recognition to achieve its automatic analysis. All studies focus on the key techniques of automatic recognition and classification for blood cells such as segmentation of blood cells, feature extraction and the classification technique. The purpose of these works is to improve the correctness of the automatic recognition and classification of blood cells. The main contents are listed below:(1) Automatic image segmentation is the key step since the results of segmentation directly influences the subsequent feature extraction and recognition. It is possible to get better segmentation results to adopt a fixed threshold value in color space when the stain is good, but it is difficult to satisfy this condition. So an automated segmentation algorithm fusing gray space, colorful information and mathematical morphological gradient is proposed for segmentation of the nucleated hematopoietic cells (including nucleus and cytoplasm). For the accurate segmentation of the nucleus, the conventional iterative threshold segmentation are improved. Color information and prior knowledge are fully used by transaction of color spaces. In order to prevent over-segmentation, the morphological gradient information is used to mark the background, nucleus and cytoplasm. The edge detection is implemented in gray gradient image since the morphological gradient can detect the contour better than other conventional edge detection operators. The results show that the method is valid and efficient to segment color images from blood smears.(2) The characteristics of blood cells are the basis for their classification. After the cell segmentation from the complex background, the next step is to find distinguished features for analyzing and classification, that is to conduct feature extraction. Based on the segmentation images, the characteristics including morphological features, chrominance and luminance features, color features, texture features and particle features are extracted. To make full use of particle features of blood cells, a technique of circular particle detection is proposed to extract cytoplasmic granules. The simulation results show the effectiveness of this feature.(3) For the classification, in order to solve the problems of high characteristic dimension and multi-category classification, and to enhance the correct rate of classification, the application of support vector machine (SVM) is studied to achieve the automatic classification of the cells. In the thesis, the methods are presented to construct SVM and select control parameters. Meanwhile, the three-tier BP neural network classifiers is designed . Finally, the methods cross validation is used to verify the performance of SVM. The proposed method is compared with BP and the results indicates the stability performance in a limited sample classification.Finally , we make a conclusion and propose the future research directions in this field.