| With massive emergence of diverse medical imaging equipments, medical images act as indispensable tools in modern clinical diagnosis and medical research. It is important needed to be solved immediately that how to effectively organize, manage and retrieve these large-scale medical image data. With the development of Picture Archive and Communication System (PACS), many relevant problems are solved, such as the gain, display, storage, transmission and management of digital image data, etc. An important standard and protocol that ensure PACS becoming a open system is Digital Imaging and Communication in Medicine (DICOM).In the present DICOM network, image retrieval is based on text. The method of image retrieval is unable to describe the rich information contained in medical images, which confines the application of traditional text-based image retrieval technique in medical diagnosis and research. Therefore Content-Based Image Retrieval (CBIR) technology is proposed. The main idea of CBIR is to establish images'eigenvector according to their visual characters such as color, texture, shape, and spatial relations, etc. Next, to search the most similar eigenvector in the database. Finally, retrieval outcome is given by similarity. That method of retrieval generally doesn't include images'text information. But in clinical practice, retrieval need for medical images according to patients'name or age is still required.This thesis profoundly analyzes the features of medical image retrieval, and proposes a pattern of the multistage retrieval according to users'demands or goals. The retrieval pattern combines technologies of text-based retrieval and content-based retrieval. So it can be used to gradually reduce retrieval scope and get right images, in accordance with diverse search criteria.In an experiment, the first-level retrieval is based on text data. Firstly on the basis of present DICOM network for text retrieval way, change original manual input data's way, through analyzes DICOM file meta information, realize the text data automatic extraction and input,these information contain the major part information which the image involves under hospital environment, such as patient, research, sequence, image, and through these text information establish index. Users can retrieve the image through known exterior information and define the retrieval's scope, such as the image's type, the patient's age and so on, so retrieval's content to be more detailed concretely, reduce manual input work and the error-rate, enhance the input's speed of the manual data.The second-level and the third-level retrieval is based on the medical image's visual content, because the different medical image as well as the different imaging's pattern is suitable for different interest regions or eigenvectors which to extraction, for experiment the multistage retrieval's feasibility, this thesis select the mammary gland X-ray's images which have the clear classification of the pathological diagnosis to take content-based image retrieval for target. The experiment design separately aiming at the retrieval of the breast's block region and the region of interesting (ROI), firstly carries on pretreatment to mammary gland's images, through eliminate noises of the background, the muscle, the frame of the landscape orientation, then extract the statistical texture eigenvectors and the shape's eigenvectors to establish database for the region of the breast, realizes the second-level retrieval through to compare eigenvectors, and on the basis of the second retrieval's result realizes the third-level retrieval through extract the eigenvectors of the region of interesting and establishes the index, the experimental result indicate that the multistage retrieval can much more aim at the user's needs and interest information, it is advantageous to enhance the degree of satisfaction to the retrieval's result.Because based on visual content's retrieval has certain redundancy for selected eigenvectors, to obtain the better eigenvector's parameter, the experiment selects 90 illustrations to carry on principle components analysis, this way can reduce redundant information including in eigenvectors and the operand of comparison between eigenvectors, enhance the ability of the system's retrieval. Finally, the experiment give the relational graph with regard to the recall rate and the precision rate whose are about the second retrieval and the third retrieval, finally indicate that the third retrieval's result surpasses the second retrieval's, has further confirmed the feasibility and the usability of the multistage retrieval. |
PDF Full Text Request