The Study On Segmentation Technology Of Colonoscopic Images

Medical image segmentation is a very important step in medical diagnosis, analysisand therapy. Due to its complexity and variety, colonosopic images are always difficult forimage segmentation. Currently computer-aided diagnosis for colonoscopic disease has notbeen employed. Therefore, the purpose of this thesis is to provide an efficient segmentationmethod for computer-aided colonoscopic diagnosis, and then enable it to be used for imageanalysis, the recognition of colon cancer and other diseases.In this thesis, two main diseases--polypus/tumor and erosion/bleeding and theirclinical performances are discussed, and the development of medical image segmentationand several main methods are reviewed. The Random walks method is presented for thesegmentation of polypus/tumor images, by using human defined seeds and backgroundpoints; as for the erosion/bleeding images, the K-means clustering method is used forsegmentation, with human defined K and clustering in L~*a~*b~* color space.Experiments are taken on clinical colonoscopic images, and the segmentation resultsare evaluated by qualitative and quantitative analysis. Compared with other twomethods—modified geometric deformable model and Fuzzy C-connectedness, our methodis more accurate and efficient in polypus/tumor image segmentation, with the resultscompared to the Gold Standard given by doctors. The segmentation result oferosion/bleeding images by K-means clustering method is also practical and accurate.These two methods studied in this thesis will provide the future computer-aided analysissystem with a solid support for disease diagnosis.