Research On Automatic Segmentation And Recognition Of Stool Microscopy Image

In clinical pathology detection, the composition analysis of human waste is of great importance. As part of the three routine medical tests, stool microscopic examination provides an basis for computer-aided medical diagnosis. It's aimed at recognizing active substance such as blood cells, parasites and mycetes. This paper is about key technology on recognition of stool microscopic image, related to technologies such as image preprocess, image segmentation, contour search, feature detection and classification.Image preprocess includes denoising and light correction. since the backgroud of stool microscopic image is complex, with many impurities and the image background is vulnerable to the light and detection environment, therefore, the first thing is denoising to decrease grain noise. And then, in order to unify the illumination environment, a step of light correction is necessary.Image segmentation is the keypoint of many projects, which are based on image recognition,its result will influence the later process. In this paper,the Chan-Vese model is chosen,which is a segmentation technology based on level set functions.It integrates global information of an image,so as to get a good segmentation result without any edge information.Contour search is used both in later image segmentation and stage of feature ietection. The accurate detection of target contour is a keypoint for later feature detection. So a first contour search is to fill the holes of a closed area as to reduce the analysis and computation of inner contours. Another contour search is to find the whole target boundary, for each target contour, analysize its feature, including area, circularity, form factor, perimeter ratio, fitting error, etc.The last step is classification. in this paper, the decision tree is used to cell classfication. there are two phases in objective classfication. Firstly, train the standard images according to human marks, and build the decision tree according to the detected cell characteristics. The second step is auto-detection. In detail, decide each target's class with the decision tree trained before.Experiment results show that the image segmentation and recognition algorithm used in this paper can efficiently classify cells in microscopic images, which can be valuable in practise.