Research On Key Techniques Of Content-Based Medical Image Retrieval

There are lots of images, graphics, videos, audios and animation besides text information on Internet. It has already become impending need for retrieving the above media information quickly and exactly, so content-based multimedia retrieval emerges as the times require. Content-based medical image retrieval (CBMIR) is the application of content-based image retrieval (CBIR) in medical field, and plays an important role in clinic, teaching, research and PACS (Picture Archiving and Communication Systems), etc. The main research aim of this thesis is to integrate CBIR and medical image to provide convenient and accurate retrieval tool and assistant suggestions for radiologists on the situation that CBIR techniques are still immature.On analyzing key techniques and development trends of CBIR systematically, this thesis focuses on the following topics that are key problems of CBMIR: semantic features extraction model of medical image, integration of image content and text information, optimization of feature weights on multi-features image retrieval, relation between content-based medical image retrieval and computer aided diagnosis (CAD). These topics are studied and analyzed systematically. For detail,(1) Medical image description approach integrating low-level features and semantic featuresBased on general framework of CBIR and medical image characteristics, this thesis presents the framework of CBMIR and describes the function of each module. It can be seen from the framework that analysis-description module is the key step of CBMIR. Most of current image description approaches are based on low-level features including color, texture, shape, etc, which are quite different from human descriptionon image, so semantic features are proposed by researchers. There are three sources for getting image semantic features currently, which are based on knowledge, manual interaction and exterior information sources. Hierarchical medical image semantic features model is proposed in this thesis based on the three masterstrokes, and the semantic contents, listed from low-level to high-level, are general semantic features, doctor semantic features and object semantic features respectively. Next level includes senior semantic features than previous level, and the higher-level semantic features are based on lower-level semantic features. This model applies the DICOM standard (Digital Imaging and Communication in Medicine) and diagnosis report by doctor to CBMIR. Although these contents are text information interrelated with image, they are indispensable semantic features. After that, this thesis proposes medical image description model integrating low-level features and semantic features, and presents the definition of the model. Brain CT image retrieval experimental results validate the accuracy of this model, and show that the retrieval combining image contents with interrelated text information is an inevitable development direction in the future. This is the only way by which medical image database can apply to extensive application.(2) Optimizing feature weights in special image retrieval system based on multi-featuresAn image has multi-features, and different feature has multi-expression methods. How to optimize feature weights for getting better image retrieval results is an unsolved problem when multi-expression methods of multi-features are integrated. Double optimization strategy is proposed in this thesis. (1) Depending on the characteristics of special image retrieval system, the feature weights assignment is changed to optimization problem, and the definition is presented. Genetic algorithm is applied to solve the optimization problem to get better initial weights; (2) If the users aren't satisfied with the retrieval results by this optimization weights, relevance feedback is used to adjust weights dynamically. Aiming at overcoming the disadvantages of typical CBIR system MARS, this thesis proposes an improved...