The Study Of Multi-atlas Based Medical Image Segmentation

In the MR-to-TRUS guided biopsy of prostate cancer and the radical treatment,the accuracy of segmentation of MR image is very important. However, prostate MRimage automated segmentation still is a challenging work due to the weaken boundaryinformation in some areas. The expert segments these areas mainly according to theirknowledge of the anatomical location of prostate. Atlas-based segmentation, for itsfull automation and high accuracy, has become one of the popular automatedsegmentation techniques. The atlas essentially depicts the shapes and locations ofanatomical structures and together with the spatial relationships between them.Therefore, this thesis mainly studies the method of atlas-based segmentation.Multi-atlas based segmentation consists of three main steps: registration, atlasselection, and atlas combination. Registration is an image preprocessing step. Theatlas selection and combination play an important role for the performance ofsegmentation. Thus, this thesis mainly focuses on the atlas selection and combination.Specifically, in the atlas selection, it proposed a manifold learning based atlasselection, which uses the locally preserving projections (LPP) algorithm to map theimages into a low-dimensional space and selects the atlases in this space. After that,by analyzing the reason of the error of atlas selection, the neighbor tissues maymislead the atlas selection. This thesis tries to address this problem by proposing alabel image constrained atlas selection (LICAS) method, which exploits the labelimages to constrain the manifold projection of raw images. Analyzing the data pointdistribution of the selected atlases in the manifold subspace, a novel weightcomputation method for atlas combination is proposed.The main contributions of this thesis are three-fold:1) It proposes a LPP basedatlas section method to improve the performance.2) A data-driven new manifoldprojection method is developed by taking the label image information into account forselecting atlases on a lower-dimensional manifold for image segmentation.3) Theatlas combination weights are computed by solving a problem of reconstruction of data points in the manifold subspace. Compared with other recent methods, it hasbeen shown that our method is efficient and superior in performance.