Research On Shape Feature Extraction And Classification Of Red Blood Cell Image

Erythrocytes (Red Blood Cells) are produced in the bone marrow, and (after about120days) are degraded in the spleen and liver. They are the most common (>99%)blood cells and of nucleus and organelles. It is common knowledge that erythrocytes areimportant to maintain human’s normal physiology function. They deliver Oxygen intovarious organs and transport carbon dioxide produced by cell metabolism back to lung.Furthermore, there is considerable evidence authenticating that erythrocytedeformability is an important determinant in the filerability of blood and consequentlyin the pathology of a variety of blood-related diseases. In addition, the shape of redblood cell is also a determining factor for its deformability and filerability. In a matter offact, Myalgic Encephalomyelitis (ME) and Multiple Sclerosis （MS） arise from thedegradation of erythorocyte deformability in pathology research. Therefore when clinicdiagonosing, the shape analyzing of erythrocyte is helpful for physician to discriminatewhat kind of state of an illness the patient suffered.The most mainly used method for analyzing erythrocytes deformability iscomputing the shape distribution of various kinds of cells in a red blood cell image. Dueto the variability of erythrocyte shape, it is not appropriate to employ traditionalmethods here like2D image processing technique any more. The difficulty of adoptingthe existing results of former studies from the laboratory to clinical diagnosis ormeasuring the effectiveness of treatment is mostly on account of the large amount oftedious work required to measure the distribution of different erythrocyte cells from asample. It is time-consuming and low efficiency when using traditonal artificialanalyzing, which rely a lot on human’s intuitive impression estimation. Accordingly,the erythrocytes’deformability and shape information should be taken intoconsideration to build up an automatic processing system.The original input red blood cell image we are aiming to deal with is captured byScanned Electronic Microscope (SEM). We map the intensity level image into a3-Ddepth through Shape from Shading (SFS) first to get the height field of each pixel, theneach cell on top level is extracted from the whole image by means of region growingalgorithm based on boundary contour tracing, which are used for feature extraction andstatistics computation. Finally an automatic classification and recognition system isimplemented in the end. The detailed information of each chapeter is decribed as below. 1. Due to high quality of red blood cell image captured by SEM, which present clearshading information, we proposed a three dimension reconstruction approach for cellsurface shape based on seeking a solution for the reflection equation solving usinglinear approximation. We made an appropriate assumption that the cell surfacereflection is Lambertian model, linear approxipation difference based on Taylorexpansion is employed to solve the Image Iradiance partial diffenrential Equation (IRE).The reconstructed three dimension height field can be viewed as reange image which isto be input for cell image surface segmentation.2. As there are lots of overlapped red blood cells in the original image, we onlyinterested in the top level cells with regarding to statistics feature. Each cell is extractedindividually from the whole image by using region growing algorithm which is basedon boundary contour tracing. The segmentation algorithm consists of two mainprocedure, contour tracing and region growing respectively.3. We put forward an adaptive curved surface fitting and curvature computationmethods for irregular cell shape with multi-deformation. After height field reconstructedwith Shape-from-Shading technique, the three dimensional data points are used forsurface fitting based on least square method. The fitting points are selected in terms ofdepth root mean square error (RMSE) threshold pre-set. As we well known, the surfacetype can be expressed by the sigh of Gaussian curvature and mean curvature. We adoptbi-variate polynomials functions set to realize the surface fitting from which the surfacepatches can be segmentated successfully.4. In the chapter6, we reviewed the background knowledge of support vectormachines and particle swarm optimization algorithm and a cascaded architecture ofSVM is proposed here. We introduced a PSO-CSVMs classifier for erythrocyteclassification and recognition automatically. Particle swarm optimization is used tooptimizing nonlinear kernel function parameters and selecting feature subset.