Study On Key Techniques For Content-based Medical Image Retrieval

With the prevalence of modern medical imaging equipments and the more and more attention of people to health care and healthiness, medical image processing and analysis is attached more importance. In the face of a large of medical image data from different imaging equipments in allusion to different human organs and their different parts, it is an important problem cried for solving in the field of medical image that how to effectively manage and analyze these data. The techniques of Content-based image retrieval (CBIR) provide the foundation of information analysis and the support of decision making for the management of the medical image data and Computer-aided diagnosis. However, due to the characters of amount largeness, imaging complexity, mode multiplicity and strong pertinence in medical images, there are some new opportunities and challenges for the content-based medical image retrieval (CBMIR). This dissertation focuses on CBMIR. Base on the detailed and systemic analysis of key techniques and future trends of CBIR, some exploring research works are developed in this dissertation, which include object segmentation, feature extraction, high-dimension index, relative feedback, and so on. The main works of this dissertation are summarized as follows:(1) A hybrid segmentation algorithm combining watershed transform and space preserving Fuzzy C-means (FCM) clustering is proposed in this dissertation. The image space can be partitioned into homogeneous sub-regions by watershed transform using the mathematic morphological principles. Then significative segmentation regions are gained through space preserving FCM clustering to merge watershed regions. To optimize the unsupervised clustering approach, firstly initial clustering centers are measured by the probability density distribution. Then sub-region fuzzy connectivity relations are analyzed to obtain the fuzzy partition matrix for clustering inputting. Lastly the consistency of object segmentation is preserved by introducing the space distributive features to the objective function of FCM clustering. The hybrid algorithm is applied to segmentation of Computerized Tomography (CT) images, and the experiment results show that the algorithm is more effective for medical image segmentation.(2) The methods of the visual feature extraction are proposed respectively based on textures and edges for CBMIR. Firstly a method based on co-occurrence matrix and wavelet attributes is used for image texture feature extraction in this paper, which combines the image’s statistical characters in the frequency domain and spatial distribution attributions. Secondly applying the results of object segmentation in medical images, boundary features are extracted based on shape and density histograms and shape features are described through compact degree and Fourier descriptor. Lastly the extracted texture features and edge shape features are respectively applied to medical images retrieval, and the results show them have better effect comparing to other methods.(3) A high-dimension indexing approach based on clustering analysis is proposed in this dissertation for medical image retrieval using integrative features. It is hard to avoid "dimension crisis" for integrative features extracted through different ways and means. To keep image features possessing bigger distinctness and eliminate redundancy in feature dimensions, the feature space of images need to be shrunk in the condition of protecting image semantic. In the approach, feature dimensions are first decreased in the vector space by unsupervised clustering. Then the dimension clustering algorithm is applied to create the similar index in the feature space. Under this indexing structure, irrelative images are filtered and the search range is reduced according to cluster areas. The image retrieval efficiency is improved by greatly decreasing the time of similarity computing.(4) An improved SVM relevance feedback algorithm is proposed for medical image retrieval in this dissertation. After analyzing objective function of support vector classification, by introducing the membership of uncertain data to classifications, the separating hyperplane of improved SVM is defined partial to uncertain points to decrease learning risk and to accelerate the classifier convergence. Simultaneously, new sampling algorithms are used for the relevance feedback learning through introducing an active learning scheme into images retrieval. And an improved relevance judgment model is proposed through analyzing the feedback models of users. Compared with traditional SVM, the improved SVM learning approach has higher efficiency of precision and speed in medical image retrieval.(5) The dissertation designs and implements a prototype system of CBMIR, which applies the uniform pattern to integrate a lot of classical algorithms in medical image processing and analysis, and offers the plane for new algorithms to realize and test. The system shows validity and efficiency of the theories and approaches about image segmentation, feature extraction, high-dimension indexing and relevance feedback proposed in this dissertation, and provides a basis for further in-depth research.In summary, this dissertation focuses on fundamental problems related to CBMIR, including object segmentation, feature extraction, high-dimension indexing and relevance feedback. Some algorithms and models are proposed, and a prototype system is implemented by integrating relevant resource, in which lots of theoretical analysis and experiments show that the proposed approaches are efficient and effective. The researches presented in this dissertation are hoped to accelerate the development of CBMIR and have both academic and clinic significance.