Research On Key Techniques Of Content-based Medical Image Retrieval And Semantic Modeling

Content-Based Image Retrieval (CBIR) becomes a hotspot in the medical research and application due to enormous diagnostic Radiology data (Ultrasonic, X-ray, MRI, CT, DSA and PET etc). In order to assist diagnosis and scientific research, CBIR system can rapidly search clinical medical images that have the similar pathological feature and have diagnosed definitely already. However, only the properties or the low-level visual features of images, which can't denote real medical images, are used to describe content of image in existing medical image retrieval systems. Therefore, seeking the relations between the image characters and the medical knowledge is effective way to extend the range of medical image retrieval.Based on analysis of technology of CBIR and its developing trend, the key technique such as extraction and optimization of low-level visual features of medical image, the relevance feedback based on the optimal weight, and the semantic retrieval base on machine learning are studied in this dissertation.Considering the particularity of medical image, the algorithm of low-level visual feature extraction and optimization are research in this dissertation. A local linear embedded decreasing dimension algorithm based on constraint least square method is proposed in this dissertation. And the optimum value of local reconstruction weight vector is computed with this algorithm. The experimental results indicate that the matching complexity can be reduced effectively and the redundant information can also be taken out and thereby the purpose of efficient expression of visual feature can be achieved.Considering the difference of abstractness of image feature and subjectivity of human perception, the relevance feedback method based on feature-combined are used to solve this problem in this dissertation. An initialized weight determining algorithm based on QGA is proposed to convert the computation feature weight which makes the evaluation index best to optimization problem with the powerful global optimization ability of QGA in this dissertation. Also a dynamic weight adjusting algorithm based on grey relational coefficient is proposed in this dissertation. The grey relational coefficient is used to compute the similarity of images which can estimate the relatively importance in retrieval. The experimental results indicate that a satisfactory query effect is acquired in the initial retrieval step with the proposed relevance feedback algorithm. Simultaneously the feedback circle is shortened and more accommodate the custom query requirement.According to the analysis of content of medical images, a semantic modeling method based on support vector machine (SVM) is proposed and applied to semantic modeling of ground glass symptom in lung CT image in this dissertation. Firstly, the semantic content expression is defined according to the prior knowledge. Secondly, the low-level visual features are extracted. Lastly, the machine learning method of SVM is used to construct the mapping of low-level features and high-level semantic. In addition, a preprocessing method of lung area segmentation is proposed to enhance the samples recognition ability in this dissertation. The experimental results indicate that the proposed semantic modeling method is effective and feasible.