This project aims at continued development of our HAMMER registration software, with emphasis on significant improvement of its registration speed. Image registration is necessary for automated integration and comparison of data from individuals or groups, as well as for the development of statistical atlases that reflect anatomical and functional variability within a group of individuals. Our group has previously developed a software package for deformable registration, called HAMMER. HAMMER has been disseminated to more than 250 users over 20 countries, supported by user groups, and applied to many large clinical studies involving thousands of scans. The current HAMMER algorithm takes about 1.5 hours to complete the registration of two regular size brain images in a standard PC workstation. Although its speed is not slower than many other registration algorithms with acceptable performance, it is highly attractive if the processing time can be significantly reduced while the registration performance can be still kept similar or even improved. We propose to instantly determine a close intermediate atlas for each new testing subject. Thus, the subsequent registration algorithm only needs to estimate the remaining small differences between the intermediate atlas and the new subject, which can be completed very fast and robustly. The determination of a close intermediate atlas for the new subject can be completed by using the statistical correlation learned between the training brain images and their respective deformations to the atlas.
The first aim of this project will study particular learning methods to effectively and efficiently capture this statistical correlation, in order to instantly warp the atlas close to the new subject. A manifold learning method will be used for dimensionality reduction of training images, and a support vector regression model will be used for establishing the statistical correlation between dimension-reduced samples and their respective deformation parameters. The accelerated registration algorithm will be extensively validated and compared with HAMMER by both simulated and real brain datasets in the second aim. It is expected that better performances will be achieved for individual brain registration as well as for identification of group differences between normal and abnormal groups, because of robust registration obtained after good initialization between atlas and new subject. Also, the registration speed is expected to be reduced significantly to about 10 minutes, from the current 1.5 hours, according to our preliminary results. The final developed method will be incorporated into the HAMMER software package, to be released again to the research community.
Showing the most recent 10 out of 34 publications
