Research On Segmentation And Retrieval Key Techniques For Medical Image Processing

With the popularization and development of multimedia techniques to medical research in recent years, medical image processing systems have got important applications in clinic, teaching, scientific research, medical image storage, indexing, and communications. But, the research in medical image processing is still in its early stage and not mature, which results in unperfect diagnostic effect, so much as wrong diagnosis. How to integrate key techniques of image processing into medical images seamlessly and to provide scientific, convenient, and accurate information for doctors has been the main research goal.Above these, medical image technology, which including the medical digital imaging and medical image processing and analysis, has quickly becomes the new crossed subject in recent years, which has important function. It not only can enhance the clinical iatrical level based on existing medical imaging equipments, but also can provide digital implement way and firm foundation for medical research and education, computer aided clinical operation.Supported by the National Nature Science Foundation Project of China 'Research on data segmentation and restrict fitting in digital human project' and Natural Science Foundation Project of Shandong Province 'Research on segmenting and fitting in CT image processing', this thesis analysies the application requirement and existent problems of image segmentation, image retrieval and 3-dimensional modeling that for purpose of 3-D image retrieval in medical image processing area, and conduct systemic and comprehensive research on above key techniques. The main contributions of this thesis are described as follows:(1) The context-label tree based multi-scale MRF segmentation algorithm.The process of 2- dimensional medical image segmentation belongs to the phase of preprocessing, it has important effect to medical research, clinical diagnosis, pathology analysis and treatment. For our research, medical image segmentation can automatically identify the boundary of the human organ surface during the medical image segmentation, so that it can provide important information for image retrieval and 3-dimentsion modeling. The algorithm is based on multi-scale Markov Random Field segmentation model, combining the context label tree technique for giving attention the statistical reliability of big window and local definition of small window. At the same time, it presents the combination method to multi-scale MRF and context-label tree, which utilizes the data dependency of different data blocks on the same scale, and the data inherence on different scales.This method gets desired segmentation effect. The experiments compare the traditional Canny edge detection method and our method, and compare three segmentation methods this thesis involved. The results illuminate that the context-label tree based multi-scale MRF segmentation model is better than the dendriform multi-scale MRF segmentation model, and the dendriform multi-scale MRF segmentation model is better than the fixed-scale MRF model, which is also showed by 'error partition ratio of pixel' adequately.(2) A local medical image retrieval method based on interest points and key blocks.Following the fast increasing of medical images, how to find the needed images from large database quickly and exactly has been a important and difficult problem. In clinical practices, when doctors observed some unidentified medical symptoms, he could find similar cases from medical knowledge base or digital library use the medial image retrieval technique. Thus, it can help the doctors in identifying, curing the disease.The medical image retrieval method based on interest points and key blocks is the first method this thesis presents, which obtains the interest points of one image by wavelet transformation. Then, it partitions the image to well-proportioned data blocks, and classifies these blocks into with interest point and non interest point ones. Finally, it distills the vectors of low-level features, and transforms the matching between two images to the matching between two blocks. At the time of image retrieval, the matching algorithm is presented in this thesis, it carries through the comparability measurement of those two kind of data blocks respectively, and endows different weight to the results, for the purpose of more freedom to interest points and key blocks, and implementing different demanded retrieval operation of medical images for users.Experiments illuminate that the precision and recall of medical image retrieval method based on interest points and key blocks are all better than pute interest point or key block method. Furthermore, experiment which comparing the tradional grey histogram method and this method, also shows that the latter preforms better than the former.(3) An image retrieval method based on spatial position relationship of objectsThis is the second retrieval method the thesis presents, it adapts to the inherent specialty of medical image, and make image retrieval operation can carry through calculation by arbitrary spatial distribution and attribution of image objects. This is a hiberarchy image retrieval method, which realizes the matching from the object feature to size function to spatial position relationship, designs and implements the expression and matching algorithm of object spatial relationshipl, image retrieval algorithm. Based on medical image segmentation, this method obtains the section of main objects in image, distills the features of shape and position relationship as the result of that object, calculate the similarity between two images according to the object features and implements retrieval.Two kinds of experiment that 'query by example' and 'query by object' are implemented, which compare this method to the traditional Hough transformation and Euclidean distance retrieval method, the results show that this method obtains perfect retrieval precision.(4) 3-Dimensional medical image modelingThe 3-dimensional image retrieval needs to reconstruct the surface by smoothly connecting different images from the human organ. Amethod to construct a virtual human model by smoothly connecting three surfaces defined in different parameter spaces is presented in this thesis. The model can magnify details in the original image, so that doctors can see the structure of 3-dimensional object. That is very important for the realistic analysis of the human organ and disease symptoms.