Research And Implement On Techniques Of Segmentation-Based Medical Image Retrieval

IV With the development of multimodality medical imaging equipments and the improvement of image-based clinical technique, large amounts of digital images are produced everyday. This cause the medical images have become indispensable to modern clinical diagnosis and medical research. It becomes an urgent problem to be solved how to organize, manage and index medical images in so large scale and how to use those images in Clinical Diagnose Process (CDP).The traditional information retrieval techniques, which are based on text comparison, are not satisfied to retrieve large scale medical image databases. New image retrieval methods should be explored. CBIR provides the ability to access images by their content, which attempt to solve most of problems by retrieval system based on text mode, become one of most active researches in computer vision, image and video processing, Data Mining and so on. It is a very promising idea to introduce content-based image retrieval (CBIR) technique into indexing medical image databases.Currently the researchers have already designed many content-based image retrieval systems. While these researches establish the basis of CBIR, the retrieval performance is still far from users'expectations. The main reason is the traditional content-based image retrieval techniques always use the global feature of the image to be the index. It ignores the local feature in the image content. But in the domain of medical image process, the feature of the image we are most concerned about is no other than local feature. This paper brings forward the segment-based image retrieval system (SBIR). Segment-based image retrieval systems aim to improving the performance of content-based search by segmentation each image into a set of"homogeneous"regions and extract local feature in each region. We use these local features to describe the image content, which supply the gap in global V feature based image retrieval. Simultaneity, this paper also use the technology of multi-orientation and multi-scale Gabor wavelet transformation in the image retrieval to represent images at object-level, which is intended to be close to the perception of human visual system.This paper mainly studies the key technology in the segment-based image retrieval about image segmentation, feature extraction, feature matching and so on.Image segmentation is the key step in segment-based image retrieval. To transform the original image into regions which are provided with the same property through image segmentation, will make us extract local features of the image in each region. The effect of segmentation has a direct influence in the effect of retrieval and feature extraction. The time complexity of the segment algorithm is also one of the main factors to weigh the retrieval system's performance. Base on all of these image segmentation methods, this paper introduces the one based on Minimum Spanning Tree (MST) in the graph theory to be the image auto-segment algorithm. The MST is one of clustering algorithms in feature space. It defines a predicate for measuring the evidence for a boundary between two regions using a graph-based representation of the image. Based on this predicate, the process of constructing MST is just about the process of pixel clustering. After clustering, we have already got the segment regions.The algorithm maximizes the similarity inside the regions and also maximizes the difference among the regions. It can self-adapting transfer along with the changes of the image content. So it is provided with predominance in segment-criterion and segment-theory. This paper introduces the algorithm of Kruskal to implement MST and uses the method of'union by rank'and'heuristics of path compression', which makes the algorithm run in time nearly linear to complete the segment of image. In addition, the algorithm uses data structure of disjoint-set forest, which can description the image data easily and make the operation efficient and effective.We change the neighboring system to obtain better segment effect. It consists of two layers: the basic and super connections, which make the neighboring system to be self-adapting. The basic connection keeps the proximity of regions, and the super connection keeps the continuity of regions. They improve the capability of capturing global segment effect. The new neighboring system also eliminates over-segmented cause by the segmentation and decreases the number of segment regions in the old neighboring system.In this paper, we extract the significative feature of the image in the regions segment by the algorithm mentioned above. In additional, we analyze the local texture feature by the technology of multi-orientation and multi-scale Gabor wavelet transformation. We use both of the features to be the index for the image retrieval. 2-D Gabor function is locally optimal in both time domain and frequency domain, and it has excellent selection feature about orientation and frequency. In additional, 2-D Gabor wavelet has good similarity with mammal's vision system. So, it is very fit for the medical image retrieval. In this paper we calculates convolution between a bank of 2-D Gabor filters and the grey values of pixels in an image around every given position. This can link the pixels in an adjacent region together, and reflect the changes of the grey values of pixels in a local area of an image from different frequency scales and orientations. Thereby, we can get the local image feature. We also adopt FFT and IFFT to accelerate the rate of convolution and increase the speed of the system.We have developed a platform for the content-based image retrieval system. Based on the platform, we complete the algorithm of efficient graph-based image segmentation, the technology of feature extraction, feature matching and the segment-based medical image retrieval system. We have got very good retrieval result by using our system.When CBIR is applied to specific field, the most important and most difficult problem is how to use the knowledge in the field. There are so many kinds of images that we can not expect to use a single pattern to distinguish them. Every kind of images has its own characteristics, this is especially true when it was used in diagnose process. The rules we use images in diagnose process is what field knowledge means. How to employ these field knowledge in diagnose process is the key we introduce CBIR to medical field. This is the idea of semantic based image retrieval, is the heading direction of content based image retrieval, and it is the future work of this article.