Key Techiques Of Automatic Hyphae Detection On Confocal Microseopy Images Of Human Corneal

Fungal keratitis is a kind of infective corneal lesions, is a disease that usually caused by fungus. If not treated timely, it may cause very serious consequences such as corneal perforation, endophthalmitis, and so on. How to make use of the morphological differences between the hyphae images and neural images, through the computer to realize its automated classification, is an important approach to computer aided diagnosis. In this paper, through image pre-processing, feature extraction and classification process, to realize the classification of the normal nerve images and abnormal hyphae images. If the image is divided into abnormal hyphae images, then detect and identify the hyphae in the images and estimate the hyphae density, to achieve the estimation of the course of the disease.In this paper, in order to improve the accuracy of classification, according to the characteristics of the corneal confocal microscopy images, an improved feature extraction algorithm is proposed. The main work is as follows:Firstly, this paper introduces the state of the art, feature extraction methods and image classification technology are systematically expounded.Secondly, the feature extraction is the basis step to realize image classification, the selection of feature extraction algorithm will affect the accuracy of image classification directly. Considering that the textures are the most obvious features in corneal confocal microscopy images, in this paper, we focus on the research of texture analysis algorithms, and we make two improvements:one is to use the cross domain instead of the traditional square domain in LBP algorithm, another is to compute the average value of the pixels in the feature extraction domain as a new parameter in AMBP, and add a new decision condition to complete the choice of the threshold value. These improvements are verified feasibly and effectiveness through the experiments.Thirdly, the unknown categories of images can be divided into the normal nerve images and abnormal hyphae images. In this paper, support vector machine is used to achieve the classification.Lastly, line detection in hyphae images and the estimation of hyphae density are done. If the image is classified as abnormal hyphae image, the location of the hyphae is needed to be marked in the image. Considering the hyphae is displayed with highlight in the corneal confocal microscopy images, LSD algorithm is used to mark the hyphae areas in the images. Then the hyphae density is estimated.