Segmentation And Recognition Of Red Cells And White Cells In Urinary Sediment Images

Urinary sediment inspection refers to the inspection of pathologica l components inurinary. Through analysis of the components, such as red cells, white cells, epithelial cells andcast cells, test whether there is some disease or injury in the kidney and urinary system. As acommon and important clinical test, urinary sediment inspection is significant to clinicaldiagnosis, treatment detection and health screening. In this paper, combined with thecharacteristics of urinary sediment images, several effective segmentation and recognitionalgorithms are proposed by using digital image processing and pattern recognition techniquesto recognize the particles in the urinary sediment, both in theory study and practice.In the urinary sediment preprocessing, anisotropic gauss filtering method is applied.Compared with traditional filter methods, the method denoises images and protects edgeinformation simultaneously. In image segmentation, three combinational segmentationalgorithm is proposed. Adaptive Canny edge detection is utilized in top layer, Otsu thresholdsegmentation in the middle, and segmentation based on gradient in the bottom, eventually getcomplete segmentation result by merging each layer’s result. For overlapping cells in thesegmentation result, a new algorithm based on recursive bottleneck is proposed in this paper.It locate the bottleneck points by bottleneck rules, and has the advantage of accurate position,strong anti-noise ability, robustness and universality. In feature extraction, combined with thecharacteristics of red cells and white cells in urinary sediment,12different classes features areextracted respectively from shape, statistical and texture features, and validity experiments arecarried simultaneously. As the experiment results demonstrated, the extracted features hasstrong ability to distinguish, describe comprehensively and calculate simply. In therecognition, studied the LIBSVM software, combined with the distribution characteristics ofdata set, choose one-against-rest classification methods and RBF kernel function to constructthe SVM classifier. Using cross validation and grid search method to select the correspondingparameters, and finally designed two level SVM classifier. Finally, compared SVM results with BP neural network classifier results, demonstrated the effectiveness and superiority ofSVM, and achieved91.3%recognition accuracy.