Study On Segmentation Method Of Discrete Point-like Nerve Bundles From Nerve Slice Image

The treatment of peripheral nerve (PN) injury is an important research and clinic topic in the area of clinical surgery. Accurate anastomose of the same functional fascicular groups inside the impaired nerve is the goal of effective treatment of PN injury. To fulfill such goal, three-dimensional visualization of PN is of great importance. At present, three-dimensional visualization of PN reconstruction mainly consists of five steps, including preparation of PN slice image, image registration, edge contour segmentation of PN bundles, type recognition of PN bundles and three-dimensional reconstruction of PN. This thesis proposes an algorithm based on the Gabor filter and rough k-means algorithm to extract the PN bundles region from most of the nerve slice images. Being able to automatically extract the PN bundles region, this algorithm lays a foundation for realizing three-dimensional visualization of PNs.On the basis of analyzing characteristics of the PN slice image, the use of Gabor filter and the Rough k-means algorithm for the process of the segmentation of PN bundles is studed deeply by this thesis.This work consists of four parts, as follows.Firstly, we introduce the background and signification of the study by surveying the previous methods on extracting the PN bundles region from the nerve slice images, techniques of texture image analysis and those of image segmentation based on rough set theory.Secondly, on the basis of analyzing the PN slice image with the characteristics of low color contrast, nerve bundles are discrete point distribution and edges of the nerve bundles are fuzzy and discrete, the PN slice images are processed as the natural texture images by this paper.we select the Gabor filter to extract the texture features of the PN slice images after comparing it with the autocorrelation method, Fourier spectrum analysis and the method based on Gray Level Co-occurrence Matrix (GLCM). Constructed from the extracted texture features, target texture feature matrix (TTFM) provides valid image data for the following clustering procedure.Thirdly, we employ a rough k-means algorithm that is free of manual parameter setting for fine segmentation and the clustering of TTFM. As a result, the pixels within the region of the nerve bundles are clustered into a single category, thus making it possible to obtain the desired edge of the nerve bundles region.Fourthly, we apply the aforementioned algorithm to process different kinds of PN slice images and compare its performance with the GLCM based method. Moreover, to further verify the superiority of our algorithm, we develop another algorithm by replacing the rough k-means clustering module with a fuzzy c-means one. This algorithm is also compared with the rough k-means version. Experiment results allow us to assert the following findings.(1) In terms of the applicability of the algorithm, the new algorithm combining the Gabor filter and rough k-means clustering not only extracts the nerve bundles region fast and accurately, but also has perfect generalization performance in the sense that the parameters need not to be adjusted when processing another image.(2) In terms of txture feature description methods of slice images, the GLCM and Gabor filter are contrasted by this paper. In the experimental conditions of the same hardware and software platform and the same operating system, using the GLCM and Gabor filter to get the texture feature of3rd,7th,12nd and32nd selected randomly PN slice images with the resolution of16million pixels, then, using Rough k-means algorithm to cluster the texture matrix. The average time of processing single PN slice image show that the efficiency of the algorithm, which uses the GLCM describing the texture feature of the PN slice images and uses Rough k-means algorithm clustering the texture matrix, is about45percent of the efficiency of the algorithm used in this paper. It also can be seen from the comparison of the extracted nerve bundles region images that the algorithm used in this paper is more accurate than the above algorithm.(3) In terms of clustering algorithms of processing PN slice images,Fuzzy c-mean algorithm and Rough k-means algorithm are contrasted by this paper. Under the same experimental platform, using Gabor filter to get the texture feature of3rd,7th,12nd and32nd selected randomly PN slice images with the resolution of16million pixels, then, using Fuzzy c-mean algorithm and Rough k-means algorithm to cluster the texture matrix. The average time of processing single PN slice image show that the efficiency of the algorithm, which uses Gabor filter describing the texture feature of the PN slice images and uses Fuzzy c-mean algorithm clustering the texture matrix, is about75percent of the efficiency of the algorithm used in this paper. The comparison of the extracted nerve bundles region images show that the above algorithm can not classify impurities and nerve bundles well, so, the extracted nerve bundles region images contains both of nerve bundles and impurities. Experiments show that the clustering accuracy of Fuzzy c-mean is less than the clustering accuracy of Rough k-means in terms of processing PN slice images.Finally, this paper summarizes the work carried out on this paper, and gives the recommendation of the follow-up study.