Medical images are the important basis of modern medical research and disease diagnosis. Morphometric analysis aims to detect and quantify more objective and more accurate morphological features that are related disease and other factors. Morphological features are mainly used for the research of human physiological mechanism and for the computer-aided diagnosis. Existing morphometry methods mostly assume one-to-one correspondence between images in order to identify correspondences of a morphological unit in all samples. This assumption ignores the variety of morphological patterns presented in diverse anatomical structures, resulting in inaccurate correspondences between images. The diversity problem has reduced the accuracy of morphometric analysis, and thus restricting the application of the morphological features in medical research and disease diagnosis. This thesis focuses on the morphometric analysis problem of diverse anatomical structures. Firstly, for the pattern representation of diverse anatomical structures, the thesis studies the representation technology with local features of three-dimensional images. The spatial location, scale and regional texture information of local features are used to distinguish the patterns which are caused by the diversity. Secondly, considering the data characteristics of three-dimensional images and the demands of morphometric analysis, the thesis studies the improved local feature extraction algorithms, in order to enhance the level of detail and distinctiveness of local features for the characterization of localized patterns of three-dimensional images. Finally, the thesis studies the morphometry methods for diverse anatomical structures according to the accurate and effective correspondences, in order to improve the accuracy and interpretability of morphometry results. The morphological features are then combined with pattern classification technology for research of disease diagnosis. The main work and contributions are as follows: (1) In the aspect of pattern representation theory for the diverse anatomical structures, stable template patterns are constructed according to spatial distributions of local features in three-dimensional images. Correspondences between images are identified according to the distinctiveness of template patterns, therefore overcoming the inaccuracy problem from diversity. The measurement criteria for the effectiveness and accuracy of the correspondences are proposed to quantify ...
修改评论