Algorithm For Automatic Detection Of Colon Cancer

In recent years, the morbidity of colorectal cancer grows rapidly. To improve the survival rate, early precise cancer diagnosis becomes a necessity. Currently, colon cancer histopathological diagnosis is the most reliable method. In order to lighten the pathologist’s work intensity and improve the accuracy of diagnosis, computer-aided diagnosis for colon cancer using pathological images is proposed in this paper.Combining the guidance of pathologist with the structural characteristics of pathological images, two effective methods for colorectal cancer diagnosis using pathological images are proposed. The main work is as follows:(1) Recognition algorithm for colon cancer using pathological images based on low level features is proposed: improved GLRLM combined with HOG(Histogram of Oriented Gradient) are used for recognizing cancerous images. On the one hand, Color image was quantized into three colors through K-means by making full use of the color features and then the run length texture features are calculated, overcoming the disadvantages of the traditional GLRLM, which directly quantizes it into fixed grayscale gradations without fully considering the characteristics of the image. On the other hand, the gradient of some images are more obvious. Thus, HOG and improved GLRLM are combined. Then, mRMR for feature selection and SVM for test are executed. Experimental results show that the accuracy of combined algorithm is higher than the single improved GLRLM algorithm.(2) Recognition algorithm for colon cancer using pathological images is proposed by combining statistical characteristics and Delaunay features based on object-oriented method:In order to take the background knowledge and the content of image into account. First, PCA-KMEANS is used for preprocessing. then, Heuristic search is carried out for segmentation, In this procedure three objects are segmented(lumen with epithelial cell cytoplasm、nucleus、stroma). After that, features based on statistics and Delaunay are extracted. Finally, mRMR method for feature selection and SVM for classification are carried out. Experimental results show that the pathological features gain higher accuracy compared with low level features.