Automatic Algorithm Design And Implementation Of Identification Of Medical Endoscope Images For Abnormal Lesions

In recent years, increasing work pressures and rapid life pace have led to a series of psychological and physiological problems. Due to these factors, incidence of gastrointestinal (GI) tract disease is increasing. GI tract cancer has been one of the leading causes of deaths. Studies have shown that early detection and diagnosis can greatly reduce the possibility of neoplasia of the digestive tract mucosa and mortality, otherwise, the rate of patient survival is less than10%five years. In order to overcome the physical limitations of the traditional optical fiber endoscope, Wireless capsule endoscopy (WCE) is developed, which has been authorized accesses to markets by the U.S. Food and Drug Administration (FDA) in2001.However, analyzing massive images WCE produced each time is tedious and time consuming to physicians. So it is necessary to develop an effective computer-aid approach to help clinicians to discriminate amongst regions of normal or abnormal tissue.In this article, considering the impact of the color space for texture extraction, we compare the five common color spaces:RGB, CIE-Lab, XYZ, HSI and K-L, and choose the one which is most suitable to descript the texture characteristics. On this basis, we use two discrete wavelet decompositions of the image which could give a multi-resolution analysis; co-occurrence matrix is employed to extract second-order statistical texture features, then covariance signatures are computed which form the ultimate texture feature vector, called color wavelet covariance texture (CWC). Accurate image segmentation and classification is achieved by a selected Texton Boosting algorithm. For each pixel in the image, the algorithm constructs a weak classifier through repeatedly selecting label. The combination of weighted confidence of those weak classifiers ultimately leads to a strong classifier on the base of conditional random field model(CRF).The entire algorithm is programmed with C#, which is built on the platform of the Microsoft Visual C++2005Professional Edition software. A large number of experiments have been carried out to analyze the color space selection and texture feature extraction of the algorithm. The experiment shows that:K-L space is the best choice among the five color spaces to describe the characteristics of texture. multi-band wavelet transform(3-band wavelet transform) is more suitable for the extraction of texture information than the binary one.3band wavelet transform combined with the co-occurrence matrix results a72-dimensional CWC texture feature vector. The classification accuracy can reach78.47%sensitivity and83.54%specificity with800rounds of texture boosting of the feature vector. Comparing CWC texture feature extraction with co-occurrence matrix texture feature extraction and chromaticity moment texture feature extraction, we come to a conclusion that the CWC texture extraction method performs superiorly over the other two methods, either in sensitivity or in specificity in texture feature extraction.