Research On Application Of Support Vector Machine In Liver B Ultrasound Images Classification

Typically, clinicians observe and judge liver B ultrasound images with naked eyes basedon their experience. It will result in missed diagnosis or misdiagnosis because of theobserver’s negligence, limited diagnosis level, and the ultrasound images’ low gray-scalecontrast which makes it more difficult for human’s visual resolution and would lead to visualfatigue. So it needs to establish an objective method to provide necessary auxiliary means fordoctors to diagnose liver disease.In recent years, in the study of liver B ultrasound images, most of the classification ofliver B ultrasound images used artificial neural networks and single feature, while singlefeature is relatively one-sided description of images. So the classification results were usuallynot good. To solve this problem, this paper used multi-feature and support vector machine(SVM) to classify liver B ultrasound images. First, features based on statistical moments,features based on gray level concurrence matrix, features based on Fourier transform andfeatures based on wavelet transform were extracted and combined to form integrated features.Finally, SVM was used to conduct classification to determine the kernel function suitable forliver B ultrasound images and the best feature combination.Traditional SVM would gain low recognition rate if its parameters selection areinappropriate. To avoid this, a method for classification of liver B ultrasound images wasproposed combining particle swarm optimization (PSO) and SVM. First, PSO algorithm wasused to optimize the parameters of SVM to get the best parameters. Then SVM based onparticle swarm optimization (PSO-SVM) with single feature and feature combinations wasused to classify the three types of liver B ultrasound images. And the PSO-SVM wascompared with SVM based on grid-search. The experimental results show that the PSO-SVM algorithm has higher classification accuracy and better performance. Particle swarmoptimization has been played a good role in the optimization of support vector machineparameters and its optimize performance is better than the grid search method’s.To measure SVM classifier’s performance, the SVM methods were compared withK-means clustering and BP neural network. Experimental results show that methods based onSVM have higher accuracy and better performance in classifying liver B ultrasound images.