Research On Automatic Localization Of Point Landmarks In Magnetic Resonance Images Of The Brain And Its Application

Localization of the point landmarks in magnetic resonance (MR) images of the brainis essential for medical image segmentation and registration, medical image analysis andneurosurgical planning.Due to the specificity, complexity of brain MR images and thediversify of the point landmarks'features, the methods of localization of the point land-marks are different.The dissertation mainly addresses the methods of automatic localization of the pointlandmarks in bran MR images and their application. It focuses on the reference sys-tems related 2D point landmarks, 3D point landmarks extraction, corresponding 3D pointlandmarks of inter-subjects'subcortical nucleus, and the target in functional stereotacticneurosurgery.The main contributions of this dissertation are summarized as follows:Talairach and B-F reference systems are two important reference systems of thebrain, which are defined based on four 2D point landmarks on the midsagittal plane, in-cluding anterior and posterior commissures, the fourth ventricular fastigial point and theintersection point between the line passing the fastigial point and the line tangential tothe fourth ventricular ?oor. In this paper, a method of Gaussian mixture model-basedsegmentation for localization of these 2D point landmarks is proposed. The method iscombined with the anatomical knowledge and neuroradiological information. Gaussianmixture model-based segmentation is used to segment the structures in the region of in-terest (ROI), i.e., the corpus callosum, fornix and the fourth ventricle. The landmarks arelocalized according to their spacial relationship with these structures. The experimentalresults illustrate that the proposed method is fast and accurate in locating these 2D pointlandmarks.The existing differential operators for localization of 3D point landmarks are sensi-tive to noise and usually extract numerous spurious landmarks. The parametric model-based approaches are not always for the limited forms of the parametric model. In this pa-per, a method of model-based partial segmentation for localization of 3D point landmarksis proposed. To determine the landmark model and its ROI, a method of point-anchoredsurface registration is proposed. A method consisting of Talairach transformation and edge-based rigid registration is used to register the model to individual MR images anddetermine the ROI of individual landmark. Level set-based segmentation is adopted toextract the structure in the ROI. A coarse-to-fine method is presented to localize the land-marks by analyzing the segmented part. The experimental results show that our methodis accurate, fast and robust in localization of 3D point landmarks of the brain.The main problem of morphometric analysis of the cerebral nucleus is localizationof corresponding point landmarks of inter-subjects'subcortical nucleus. In this paper,a method of automatic extracting corresponding landmarks of the subcortical nucleusfrom the population of segmented MR images is proposed. Firstly, the template shape isselected from all samples in the study according to overlap coefficients, and is triangulatedto extract the landmarks. Then, the iterative closed point method is improved to align thetemplate shape to the other samples and establish the coarse correspondences betweenthe template shape and the samples. An active surface model is presented to recoverthe shape of the other samples. Thus, corresponding 3D point landmarks are extractedfrom all samples. Results of the experiments on the subcortical nucleus show that thepoint distribution model constructed with the training set obtained with our method hascompatible or even better quality in terms of compactness, generality and specificity incomparison with the state-of-art SPHARM and MDL methods. While the computation ofour method is less.For the target- pedunculopontine nucleus for treating Parkinson's disease is invisiblein T1-weighted MR images, it is difficult to be localized. In this paper, a method of au-tomatic locating the neurosurgical target-pedunculopontine nucleus is proposed. Firstly,T1-weighted MR images and proton density MR images are used to create a template ofa segment of the brainstem containing the whole pedunculopontine nucleus. And then,B-F reference based global registration, intensity based local rigid registration and freeform deformation based nonrigid registration is used to register the template to the cor-responding part in the subject's MR data. The position of the pedunculopontine nucleusis mapped from the template to the subject. Results of the experiments with 11 subjectsillustrate that the accuracy of our method is acceptable.