Study On Leukocyte Classification Algorithm Contained In Imaging Flow Cytometry

Leukocyte classification is very important in clinical blood examination, for it plays an important role in blood examination and researches on blood components. Cell recognization techniques tend to gradually take the advantage of some special instruments and softwares to classify cells automatically instead of doing it by hand. During the process of software development, dot plots and histograms are widely used, but the physical and chemical characteristics are indirectly got from these images or graphics. It's may be easier for researchers and doctors to form an intuitionistic concept by extracting features directly from cell images, such as shape, texture and so on, and maybe higher precision can be reached. Imaging Flow Cytometry (IFCM) is one kind of brand-new techniques at present. Because of its speediness and precision, widely application can be expected in blood examination in the future. The analysis of cell images gained by IFCM is presented in this paper for the purpose of classifying cells quickly and exactly.There may be some noises in the images because of the influence of CCD and environment during the collection and transmission of them. Wavelet transform combined with pointwise variance method was applied to denoise them. Then after exactly analyzing the characteristics of cell images, HSV transform was used for the denoised images and OSTU2 was applied in the HSV color space for splitting images. Fuzzy C means method being used for improving the velocity of the algorithm was also presented. Feature extraction and selection was discussed after the interested regions had been extracted. Satisfactory feature selection method was applied for the consideration of time complexity and amount of information of the features. At last, the results by using ANN and support vector machine were discussed.Primary experiment through 180 cell images of three main kinds of leukocyte (neutrophile, monocyte, lymphocyte) selected from the internet indicated that the process introduced in this paper had effiently improved the velocity of classifying without neglecting the precision. By ANN, the precision was 80%-90%, and by SVM, the precision was more than 90%.