Research And Recognition Of Liver Ultrasonic Images Based On Texture Analysis And PNN

Ultrasonic imaging is a popular and non-invasive tool frequently used in the diagnosesof fatty liver, In the traditional diagnose, Clinician make a decision based onobserving images by naked eyes, so the conclusion of disease is influenced by doctors'experience and subjectivity. AN objective method should be proposed to be used to helpdoctor make a proper diagnose efficiently.The ultrasonic liver images have various granular structures described by 'texture',therefore,the analysis of ultrasonic image can be viewed as the problen of texture analysisand classification. Utilizing the predominance of the wavelet transform and artificial neuralnetwork for the recognition of images, we do some research about the classification ofnormal liver ultrasonic images and fatty liver ultrasonic images.In the part of image pre-processing ,first reducing noise of imgages based onwavelet,then enhancing contrast of it.In the part of extracting texture feature,we analysis ultrasonic liver images based ontexture feature in spacial and frequency domain.In spacial domain ,several texture statistical anylysis techniques such as the statisticalfeature based on grayscale historigram, the gray level difference statistics and spacial grayco-occurrence matix,are thoroughly studied and actural texture meaning of every statisticalfeatures is qualitatively analyzed.In the transform-based method,wavelet texture analysis is thoughly studied. Firstly ,the basic knowledge about wavelet transform is briefly introduced, Then the wavelettransform is used in analysising images based on multi-resolution analysis.WT are used toanalyze the texture image. The low frequency coefficients denote the approximation of theimage, and the high frequency coefficients denote the details of the image.So the characterdeduced from WT coefficients can embody texture feature of images. The energycharactorization and variance characterization of sub-image are extracted.In the last part,neural netwoeks are employed to classify patters based on learningfrom examples.Because Back Propagation Neural Network(BPNN) has low convergencespeed,susceptibility to the local minimum. The RBFNN can avoid these shortcomings . TheProbabilistic Neural Network developes from RBFNN and approaches the Bayes theoremmaximum posterior probability,thus it is employed as a classifier in our method. then theabove statistical features are applied for texture classification by probabilistic neuralnetwork. Experimental results show that statistical features extraced from WT coefficientsare superior to others,and achieves good effects.