Content-based Ultrasonic Medical Image Retrieval

There is an enormous need for visual information organization management, andretrieval in the growing field of digital archives or electronic patient recorder and bythe increasing application domains of medical imaging and PACS(Picture Archivingand Communication Systems). In particular, techniques of content-based imageretrieval(CBIR)have been major topics of research for medical image databasequeries instead of text-based searching techniques in recent years.In this study, the applications of CBIR in medical image are summarized; keytechniques of CBIR are analyzed, including feature extraction, similarity andevaluation. The main content of this thesis is as follows:1. In order to eliminate character information around image, entropy and regiongrowing are utilized to preprocess medical ultrasonic image.2. Color features (global histogram, block histogram, banded auto-correlogram,Legendre moment, block Legendre moment) and texture feature (gray levelco-occurrence, Harr wavelet, Gabor wavelet) of medical ultrasonic image areextracted. Retrieval based on various features has been implemented, and theresults have been compared.3. On analyzing various retrieval results, a new method of retrieval has beenproposed, that is hierarchical retrieval—the first retrieval is performed usingtexture feature, the second retrieval is carded out using color feature.4. The gray ratio of near field and far field of ultrasonic medical image is proposedto retrieve severe fatty liver.5. Content-based medical ultrasonic image system platform has been established,which meets the need of subsequence exploitation work.Based on the experiments with image database, the retrieval efficiency usingglobal histogram, banded auto-correlogram, Legendre moment, block histogram,block Legendre moment, gray level co-occurrence, Harr wavelet and Gabor waveletare respectively 75.2％, 70.4％, 77.2％, 78.3％, 79.8％, 60.1％, 64.6％and 68.2％, theefficiency of hierarchical retrieval is 83.2％. It can be concluded that block Legendremoment is better in color features and Gabor wavelet in texture feature. So,hierarchical retrieval with block Legendre moment and Gabor wavelet improvedretrieval efficiency and it's better than weighted average retrieval.