Research On Semantic Retrieval Technology Of Mammogram

The rapid development of computer and internet techniques, the storage and conveying images turn more and more easy, results in large quantity of image information. Recent years, many research focus on content-based image retrieval. However, medical images are different from common images, the content of medical images are objective and stable, containing semantic information involving pathology diagnosis information, the research of medical image semantic retrieval is settled urgently .The dissertation applies semantic retrieval to mammogram, the mammography is the important for detecting and diagnosing breast cancer early. The tiny calcifications on mammogram is the only judgment to diagnose breast cancer, but the detection is still a difficult task, if through the retrieval of results of past case , could help doctor to diagnose effectively.The main study of the dissertation is as follows:Firstly, it is proposed to use mixture Bayesian network to model micro-calcifications semantics. The micro-calcifications on mammography are located, using both the morphology and the Wavelet Transform methods to realize it, and then grayscale, textural, morphological and calcification cluster feature of micro-calcifications are extracted. Classifying these features by SVM, which aim to extract middle-level semantics, and fusing the feature semantics and semantic defined by doctor's diagnosis into Bayesian network, finally the high-level semantic through the BN inference are obtained, which are expressed by probabilistic values. Meanwhile, every mammography was annotated by semantics.Secondly, it is the mammogram semantic retrieval. When retrieving images based on query image, a mixture methods is proposed, which combining semantic based image retrieval with content based image retrieval. First, N images are returned according to the multi-features distance between image in database and query image by using content based image retrieval method; then the first retrieval results are reordered according to semantic features. The semantic similarity measurement is a crucial step in semantic retrieval, a hierarchical semantic similarity measure is proposed, for high-level semantic, employ distance in posterior probability space, and for middle-level semantic, employ Euclid distance.The purpose of medical image retrieval is to assist doctors to diagnose cases, the research on hierarchical content of mammogram, knowledge structure expression and auto extraction of semantic feature in the thesis, provide some reference.