Research On Medical Image Retrieval Technology Based On Relevance Feedback

The current image retrieval systems are almost based on content, and facing the mainproblem of “semantic gap” between low level features and high level semantic. So therelevance feedback technology is used to solve this problem. However, in the medical imagesretrieval, doctors often ask not only more relevant images, but also to separate the results intodifferent categories.Firstly, in this paper, we propose a medical image retrieval system based on relevancefeedback framework. In the framework, Region of Interest (ROI) is extracted in thepreprocessing as the semantic information of medical images, and then the Genetic Algorithmis designed for ROI clustering. According to user’s feedback information, the Diverse Densityalgorithm proposed in the Multiple Instance Learning Framework is adopted to capture user’sreal intention and realize effectively medical image relevance. Experimental results show thatthe proposed algorithm has higher precision and recall ratio and less time.Secondly, in this paper, a tree classification structure based on SVM is proposed whichcan be used to divide medical images into three categories: normal image, the illness in leftbrain image, the illness in right brain image. The classification algorithm can also solve theproblem of less samples of user feedback and sample asymmetry problem. Experimentalresults show that the proposed classification algorithm has high accuracy.