Medical Image Retrieval Based On Texture Features And BP Neural Network

Content based image retrieval (CBIR) is a method to retrieve the similar images bythe image content. It has been widely used in the field of medical image retrieval.Because of the characteristics of the medical images and the requiring of classificationaccuracy, the retrieval of such images for doctor’s diagnosis becomes more difficultthan the retrieval of natural images. How to classify the medical images into differentclasses is an effectively way to solve this problem.Based on the introduction of the low-level visual features, relevance feedbackmethods and performance evaluation criterion, a novel method based on the texture andBP neural network technology was proposed in the paper for medial imagesclassification and retrieval. The main research and innovation of this thesis are asfollows:(1) The texture features description methods were summarized firstly. Then, animproved texture feature extraction algorithm was proposed based on the analysis oflocal binary pattern (LBP) and center-symmetric local binary pattern (CS-LBP). Theeffect of the center pixel to the local texture features was considered in the new method.Experimental results show that the algorithm has better retrieval performance for textureimages and medical images retrieval.(2) On the basis of BP neural network standard BP algorithm, a new method wasgiven based on additional momentum and adaptive adjustment of learning rate.Experimental results show that this algorithm has better classification identificationability than the traditional BP neural network.(3) The improved texture spectrum descriptor and improved BP algorithm wereused together for medical images classification and retrieval. Experimental results showthat the new method improves the efficiency of the retrieval comparing with usingdirectly texture features.