Multistage Medical Image Retrieval System Based On Multi-features

With the development of medical imaging technology and the popularization of hospital information network, the number of medical images that can be used for clinic, teaching and research are rapidly increasing. Because content-based medical image retrieval (CBMIR) system can effectively manage and retrieve a great quantity of medical image data, it has attracted more attention in the medical image processing community.In recent years, researchers at home and abroad have done plenty of researches related to CBMIR and obtained some achievements, but deficiencies still exist, such as poor portability, simplex retrieval mode, difference from human beings subjective judgment and so on. How to expand the application range of CBMIR, meet the various demands for users and supply doctors and scientific researchers with convenient, fast and accurate retrieval images is the key of medical image retrieval technology in the future. This article is about the key technology of CBMIR, compared and analyzed with the feature extraction methods detailedly according to the characteristics of medical images in various types. On the basis of constructing medical image database, a multistage and multi-features medical image retrieval system has been designed. The main research work is showed as follows:①To extract text information from DICOM images automatically: The inclusion of semantic text related to the imaging modality and patients would help to reduce retrieval range and time cost. Based on the standard data model and format defined for DICOM images, text information within patient, study, series and image in DICOM files was extracted and indexed automatically for integrative retrieval with content features extracted below.②To describe visual contents of images use multi-features combination method: According to the characteristics of low level features, different feature combinations, which combined texture features with shape features, global features with regional features, were designed for various retrieval objects, in order to make up the deficiency of single feature.③To design a multistage retrieval mode for different images: The first-stage retrieval is used to pre-screen images by matching DICOM text information, the second-stage retrieval is based on global images, by which users can retrieve similar images in overall vision and reduce retrieval range for subsequent ROI retrieval, the third-stage is used to retrieve the similar lesion areas that doctors pay more attention to.④To explore three-dimensional (3D) medical image retrieval technology: We use 3D GLCM, Shape Index (SI) and Curvedness (CV) to express characteristics of 3D colon polypus images, and then set up index in order to offer users a 3D retrieval means that can reflect lesion more comprehensively. Validation experiments using a testing image database composed of CT colon images and MR bladder images indicate that, the integration of multiple features, especially that of morphological features with texture feature, could express visual contents of medical images more effectively and the retrieval performance using multiple features outperforms those only using single type of features. Compared to one-stage retrieval, the proposed multistage retrieval could reduce the retrieval range according to users different demands, and demonstrate higher retrieval accuracy. Moreover, experimental results show that since 3D features reflect the shape and properties of ROI more specifically, the use of 3D ROI features would further improve the retrieval performance compared with 2D features, indicating its great potential in clinical detection and diagnosis. In summary, multistage and medical image retrieval system based on multi-features not only combines text-based with content-based retrieval, but also integrate local-based ROI features with global image feature, as well as 2D features with 3D features. Therefore, it could improve the accuracy rate of medical image retrieval effectively and its feasibility has been validated by preliminary experiments.