Content-Based Medical Image Retrieval

With the development of multimedia and Internet, more and more images are available. It lias become apparent that new tools for organizing these images should be developed for efficient storage and retrieval. Interest in research into content-based image retrieval has grown rapidly over the last few years. The research in this paper mainly focuses on applying CBIR methods into medical image retrieval.The key point of CBIR is to select and extract proper features. Some features which are commonly used in CBIR only include global information in the image. So new methods are proposed to segment the image at the first step and the values of these features in each small region are calculated after that. By doing so, these features can contain both global and local information in the image.In this paper, a method which segment an image into 4X4 rectangular regions is proposed. Some features are then used to describe the information of each block, such as moment invariants, global histogram, local histogram, texture and shape.Since to segment an image into 4X4 rectangular regions is not a very reasonable method, a new method is applied in this paper to segment the image automatically. The Competitive Hopfield Neural Network is used to segment the image into several classes. Then the main class features and the main block features of the image are extracted which can also be used for image retrieval.In order to measure the efficiency of all these features, all the features are tested individually to retrieve the image database. A weight is given to each feature according to the efficiency of each one. After that, all these features are combined together and a better result is obtained under this situation. A feedback approach is also proposed to refine the query representation (weight updates) to better suit users' need.