The Research And Implementation Of Three-dimensional Ultrasound Image Segmentation Methods

As ultrasonic dia gnosis is real-time, easy to use, noninvas ive and safe, it becomesan irreplaceable part of clinical dia gnosis in modern society. Tissue lesions of humanbody may bring huge bad influence to people’s daily life. It may become heavy diseaseif it is not diagnosed or treated in time, which is a threat to life. It is in urgent need toinspect and diagnose the diseased tissue effective ly. Tradit ional ultrasound imagingtechnology is only able to reappear the whole structure of the whole tissue. It can notdisplay the interested tissue lesion separately, which has been effectively solved bysegmentation technology of three dimens iona l ultrasound images.The threedimensiona l segmentation techno logy can show the three dimens ional contour of thetissue via three dimens ional display techno logy. It can show the particular interestedregion by segmentation technology as well, which is facilitated to subsequent diagnosisand analysis of the tissue lesion.There are many classical segmentation methods, such as region growing basedmethod, thresholding based method, egde detecting method, clustering method andSnake based method, etc. These methods have simp le structure and are easy to berealized. But almost all o f the m are not robust to noises and they are sensitive to theinitia l parameters and the initial model. The graph theory based method with simplestructure has rich theoretical support, which is rob ust to noise. Thus we expand thegraph-based segmentation method in2D images to3D images, which is na medgraph-based segmentation method for3D images.The graph-based segmentation method consists of3steps, which are specklereduction, graph construction and the mergence of homogenous neighboring regions. Apreprocessing procedure for speckle reduction is needed to improve the graph-basedmethod to be more robust to noises. A bilateral filtering model which has been provedto be high-efficienc y and high-accuracy especially for the speckle reduction of USimage is used as a preprocessing method in this study. The key point for graphconstruction is to generate a graph from an image. The procedure of the mergence is tomerge the vertices in a graph according to a criterion, which is called pairwise regioncomparison predicate.Ten sets of resolution phantom data,10sets of fetal phantom data and10sets ofhuman fingers data have been tested by three different methods in our paper, which aregraph-based method, fuzzy C means (FCM) and deformable model (Snake) based method. In comparison with the FCM and Snake methods, our graph-basedsegmentation method is proved to be less computationally complex, hence resulting inless computation time. In terms of segmentation accuracy, the experimenta l resultsdemonstrate that our method outperforms the3D Snake and FCM clustering methods.The average segmentation quality rate is86.31%for graph-based method,55.59%forFCM and72.58%for Snake.Our graph-based methods is robust to noises and itperformances better than the other two methods. All of the above indicates that theproposed method can improve the performance of clinical applications.