Content-based Medical Image Retrieval Using Gabor Wavelet Texture Features

With development of computer and medical imaging technology, the number of medical image data is increasing larger and larger. At present, the retrieval modes used by PACS (picture archiving and communication system) is still text-based, which cannot be able to meet the retrieval requirement of increasing and complicate image data. Aiming at resolving the problems in text-based medical image retrieval, content-based medical image retrieval has emerged. This thesis studies a kind of key technology in content-based image retrieval (CBIR), which is texture feature extracting, and investigates how to apply it into the practice of medical image retrieval.Firstly, this thesis emphatically discusses and studies several methods of texture analysis based on spectrum, including Fourier transform, Gabor transform and wavelet transform, and analyzes their advantages and disadvantages in frequency analysis.Wavelet can overcome the disadvantages of Fourier's poor ability for local analysis and Gabor's fixed "time-frequency" window, so it is particularly applicable to image texture analysis as a new time-frequency analysis tool.Secondly, the pyramid structure wavelet transform is applied to texture feature extracting of medical images.The performances of db6,db2 and haar are analyzed while extracting texture features. A new boundary extension method is introduced, which overcomes two serious drawbacks of data increasing and greater error near reconstructed signal.Finally, the Gabor wavelet texture features is applied to content-based medical image retrieval, in which fuzzy theory and significant idea are introduced. The energies in specific scale and direction are calculated using Gabor wavelet, and then the significant multi-scale and multi-direction fuzzy set is identified according to the energy. Each scale and direction of the set is quantitatively analyzed to determine its significance. The results of the analysis are introduced into the similarity measurement as weights. The introduction of significant multi-scale and multi-direction fuzzy set relieves the problem of dimension disaster to some extent. The introduction of membership into similarity measurement as weight enhances the system's retrieval performance further.