Study On Multimodal-based Retrieval Of Mammography

Medical imaging case consists of medical image and the medical diagnosticcontent. It is one of the main targets of medical retrieval that abundantinformation of medical images and the description of related cases are used toretrieve the former cases so as to provide an effective aid to doctor in diagnosis.With the rapid increasing of medical imaging data, the medical cases of theefficient retrieval is facing severe challenges. How to improve the retrievalperformance became the focus of study. At the same time, there is “semanticgap” exists between the low-level features and the high-level semantic features,therefore, the effective combination of medical images and the relateddiagnostic description of medical cases become the key and premise of efficientretrieval.Mammography is an important basis for early diagnosis of breast cancer, asmall and granular microcalcification is an important early manifestation ofbreast cancer. The key research point of this dissertation is breast calcifiedlesions, so as to improve the case retrieval performance of calcified lesions.This study mainly divided into three parts:1. The extraction of breast case features. Those features are: low-levelfeatures of image and the semantic features of case. Firstly, the low-level imagefeatures are extracted from mammography, such as the gray features, texturefeatures and shape features, etc. And then, support vector machine is used toclassify some of the characteristics to get the middle-level semantic features,and the semantics of these characteristics are associated with the relateddescription of cases, so as to constitute a case of multi-level breast semanticstructure. After that, the similarity matrix of the multi-level semantic structure isreferred to get the semantic features of other semantics, which is based on theknown semantics of cases. The low-level features’ semantic vector and the low-level features’ vector are integrated to compose the feature vector of breastcases.2. The calculation of similarity matrix between concepts in the multi-levelsemantic structure. Firstly, a multi-level semantic model is built, the similarityof the nodes is calculated based on the tree structure; Then, Bayesian network isused to obtain the conditional probability and posterior probability betweennodes, as the positive factor and the reverse factor, weighted the similarity of thetraditional measurement, and then to get the similarity matrix of the multi-levelsemantic structure, in order to solve the asymmetric semantic similarity and thedifferent semantic relevance between child nodes and parent node.3. Multimodal-based Retrieval of Breast Cases. SQL Server2000is used tostore data, which contains the mammography, the related descriptionsof cases,the low-level features, the visual feature semantics of breast cases and thesemantic similarity matrix. VC++is used to design the interface between peopleand machine, and program with Matlab, so as to achieve the Multimodal-basedretrieval of breast cases.To demonstrate the efficient performance of this retrieval method,comparsion was done in each part of the experiment, and experimental resultsshow that the method has good validity, and improve the retrieval performance.