Research On Key Technology Of Medical Image Retrieval Based On Semantic

As the medical imaging data quantity grows larger,the information storage and management have become a burning issue in medical imaging management.In the present,the relatively mature text retrieval is widely used in Medical Image Management System(MIMS),however,it is difficult to adapt the massive growth of medical imaging data.The semantic image retrieval is a research direction to solve these problems.There are inter-relationships among different levels of medical image.Thus,we can fully use prior knowledge to build semantic data models through machine learning,in order to retrieve medical images associated with similar pathological features from medical image database.By analyzing the research on semantic retrieval of medical image,this article has run an in-depth study on some key techniques,including expression,extraction and retrieval of semantic image.Its main ideas are as follows:To extract the region of interests(ROIs)of chest CT image,this article proposed an image boundary tracking algorithm based on an improved GVF-Snake model,and appliedit for lung ROIS extraction.Adaptive mesh processing techniques has been used to optimize the snake points in this algorithm,which adopped snake points homogenization to solve the local concave contour approximation problem,while with the strategy of Gauss to control the search range dynamically in the iterative process so as to effectly approximate the real countour.The experimental results indicated that the algorithm proposed in this study is able to segment the lung area into parenchyma,thorax and mediastinum with high efficiency and accuracy.To describe the CT image features comprehensively,this article expressed lower-level vision features of these images by grey-histogram statistical characteristics,gray level co-occurrence matrix texture feature,Tamura texture feature,Gabor wavelet texture feature and SIFT feature.Based on the high-dimensional data in medical image feature expression,this article proposed a Supervised Locality Preserving Projections(SLPP).This algorithm aimed at considering the proximate points that affects the calculation of reconstruction weights and the range of the adjustable patameter t.In this article,we use the fomula,in which the difference between the reconstruction weights and the zero mean probability density function is a constant coefficient,to calculate the proximate points of the feture points,and use the reconstruction weights formula with sample variance to obtain the range of the adjustable parameter t,with the distance between feature points as the sample.The experimental results demonstrated that,using feature dimension reduction in image retrieval with this algorithm will improve calculation efficiency rather than making much difference.Based on analysis on chest CT image,this article proposed a prior-knowledge-based hierarchical semantic model of chest CT image,and identified the chest diseases as well as semantic description in two levels.Therefore,the hierarchical semantic mapping mechanism of medical image had been proposed,which is doing fine mapping after rough mapping of different hierarchical semantic by a multi-instance learning algorithm based on Support Vector Machine(SVM).The experimental results showed this method has great capacity for semantic mapping.The traditional machine learning requires us to study all over again when there is a different knowledge because we are not able to learn from similar fields with this method.By analyzing this deficiency,this article proposed a hybrid transfer algorithm,which is used for semantic annotations in medical image.This method has combined with advantages of instance-based and feature-based transfer,and fully uses the classifier adaption and expansion during the process of balanced transfer,which is a sparse multi-task transfer learning method integrated with instance-based transfer.The experiment has proved that this algorithm can effectively improve the accuracy of semantic annotations in medical image.