Automatic Segmentation And Recognition Algorithm Study On Urinary Sediment Images

Urine detection is the routine hospital testing project, and is of great significance to theclinical investigation, detection and diagnosis. In order to improve the automation level ofmicroscope image detection and reduce the cost of dyeing, by combining with urinarysediment images’ characteristics and using digital image processing and pattern recognitiontechnology this paper put forward a complete set of effective segmentation and recognitionalgorithm on the visible components of the low and high magnification microscope in urinarysediment images.In the part of image pre-processing, a variety of images collected are evaluated, andbetter images are enhanced by the Gauss filter and Laplasse beam sharpening. In the part ofimage segmentation, firstly the characteristic of urine sediment microscopy image is analyzed,and then different methods of segmentation are adopted in the two cases. The improved Sobeledge detection and P-tile threshold segmentation are adopted for Low magnification pictures,and computes with the enhancing two value image, finally gets the cell’s integrity outline. Theimproved Renyi entropy adaptive threshold segmentation is used for high magnificationimages. The minimization energy function to judge the optimal segmentation points andDijkstra method to determine the optimal path are used for the adhesion cells. In the featureanalysis part, it extracts15kinds of features separately from shape, statistical and texturefeatures, feature selection model is used for feature selection evaluation, and the feature set istested effectively. In the classifier design part, the decision tree classifier and SVM designmethod are used, and combined the two methods after contrasting the experimental results.Firstly decision tree was used to separate two categories of magnification. Then the red andwhite cells in the high magnification are classified with one to many SVM methods, casts andepithelial cells in the low magnification are classified with decision tree method. Comparedwith the experimental results after combined, recognition rate of the red and white cells isimproved; recognition rate of casts and epithelial cells reached the medical level:95%.