Research On The Automatic Classification And Recognition Of Lymphocyte

Leukemia is an abnormal proliferation of bone marrow and hematopoietic tissue disease,commonly known as "blood cancer".Its characterize is the leukocytes of bone marrow and other hematopoietic tissues show excessive dysplasia and differentiation of disorders.It is a blood disease with unclear pathogenic mechanism and will be seriously threaten the life health of the patients.Using computer technology and pathological diagnosis of bone marrow cells to quantitative analysis and automatic identification has an important practical value and application prospects in early screening and diagnosis of leukemia.However,the differences of producing and staining for the bone marrow smears、complexity of background、diversity and irregularity of the cells and overlapping cells as well as other issues lead to a great difficulty of the detection and identification of bone marrow cell image.Based on previous studies and combined with basic knowledge of pathological cells in bone marrow,this paper make a systematic study of cells image segmentation、calculation and selection of characteristic parameters as well as classification of diseased cells in bone marrow by the integrated use of image analysis and pattern recognition technology.Mainly involves the following aspects:(1)Design of bone marrow cells segmentation: This paper propose a K-means clustering on wavelet analysis. Firstly, the wavelet transform is used to remove defocus noise;Secondly, selecting the appropriate channel by extracting the RGB color space component histogram information;Finally, according to the multi-resolution wavelet analysis to extract G component with more prominent information of the image, so as to optimize K-means algorithm.It can improve shortcomings caused by random initialization cluster centers such as instability、non-targeted and easy to fall into local optimun and so on.Through a variety of algorithm experiments analysis in Matlab platform, the proposed method is validity and feasibility.(2)Design of cellular localization and extraction of lesions in the bone marrow:By analyzing the characteristics of diseased cells found white blood cells stained by Wright’s color contain lots of information,a algorithm based on mixed property in multiple color space is used to locate and extract the nuclei in it.Firstly,by comparing the G(green) and S(saturation) component characteristic differences in the nucleus,constructed convert image of RGB space and HSI space to highlight nucleus;Then,using Otsu threshold algorithm based on histogram information to binarization the white nucleus areas;Finally,experiments show that the algorithm can accurately locate and extract the nucleus, performance is good and stable,as well as be robustness in the light, uneven staining and complicated background.(3)Feature extraction of three types of acute leukemia(M3、M5、L1):According to cytopathology knowledge extracted shape、color and texture features of cells from the three categories diseases, a total 46 feature parameters.In order to achieve the ideal number for the classifier, we need for a preferred feature extraction.This paper used Relief F algorithm from improved Relief series algorithm to select the above parameters and obtained 16 characteristics of the average arranged by weight for preferred features extraction.(4)Design of classifier for bone marrow lesions cells:For the three types of white blood cells,this paper using BP neural network classifier to classification and identification.First, improved BP algorithm of the steepest decline and variable step is used to establish a network model;Second,using Matlab toolbox of neural networks for classifier designing including network creating、network training and network testing;Third, using cross-validation to training and testing the sample characteristics sets of three types of diseased cells,by contrasting with the traditional experimental algorithm analysis showed that improved algorithm can achieve many advantages,such as fewer learning times、faster convergence and higher accuracy recognition;The recognition rate of training and testing network will achieve to more than 85% when variable step correction parameters of the algorithm are set as reasonable numbers.