Research On Medical Image Feature Extraction And Classification

The computer aided diagnosis has drawn much attention recently, so the research onmedical images has gained some progress. Through the feature extraction and retrievalclassification, the medical images can be distinguished as normal or abnormal. Theclassification results can be used as a diagnostic reference views. Therefore, the featureextraction and retrieval classification of medical images become the focus of research on thecomputer aided diagnosis. The main object of this paper is the brain CT image. In order tojudge whether the CT image is normal or not, it needs to extract the feature data from imagesand make an analysis of them, and then design a classifier to classify them. There is a greatrelationship between the quality of classification and the selection of feature information aswell as the quality of the classifier. In connection with the feature of medical images, thispaper analyses the advantages and the disadvantages of the existing classical methods andproposes the improved algorithms for feature extraction and classification from the aspects ofthe feature extraction algorithms and the designing of the classification methods, so that theaccuracy of classification can be improved. The main work is as follows.Firstly, the study background and significance of the medical image are introduced. Thecurrent research status both at home and abroad is comprehensively presented. In addition, thecharacteristics of medical images, the basic principles of feature extraction method, thetechniques of image classification are systematically expounded, and this part also makessome image filtering, interpolation and other pre-processing operations.Secondly, the analysis of the feature information is the basis of image classification. Theextraction methods of medical image feature are related to the accuracy of imageclassification. Considering that the shapes and textures are the most obvious features, wefocuses on the selection of image shapes and textures. After the transforming of theContourlet, each sub-band energy and statistical distribution parameters can effectively reflectthe texture characteristics, regardless of the structured or random texture image. This papermakes some research based on Contourlet transformation. Because the Contourlettransformation does not have translation invariance for image multi-scale analysis, andinvariant moments have good invariance and anti-interference to the changes such as rotatingand scaling, the characteristics and shapes of the images can be effectively reflected.Therefore, this paper combines transformation features and invariant moments, and proposesa kind of extraction method based on them. Through the classification experiments of thetexture library image, its feasibility and good classification ability are well tested and theseare also applied in medical image feature extraction.Thirdly, in order to realize the classification of medical images, brain CT images aredivided into two categories——normal and abnormal by classifier. SVM is used in this paper.Considering SVM will encounter the problems of parameter setting and feature selection inthe process of classification modeling, this paper also introduces ACPSO to optimize theSVM on the basis of the existing optimization algorithm. Through the chaotic sequence andadaptive function to improve the problem of particle swarm premature convergence, it is easy to fall into the local extreme value, and at the same time adjusting inertia weight willaccelerate the convergence speed. Simulation experiments demonstrate the feasibility of thealgorithm and the satisfactory results in brain CT image classifications.