In this thesis, we present two novel methods for medical image volume segmentation and surface registration. The volume segmentation is conceptually formulated as a problem of clustering feature vectors representing each voxel. Feature patterns are constructed by extracting texture measures and multiscale parameters for each voxel. These feature vectors are then projected onto their leading principal axes found by using principal components analysis (PCA). The number of principal components is selected dynamically using genetic algorithms (GAs). This step provides an effective basis for feature extraction. The reduced patterns are then clustered to different, spatially connected regions using a novel adaptive connectivity satisfaction self-organizing feature map (CSSOFM). This network, which is a type of Kohonen feature map, combines clustering and labeling in one network. Topological constraints are imposed on the clustering algorithm so that only voxels that are connected to each other are grouped together in a certain class. The choice of the optimum number of classes is performed automatically by maximizing a segmentation quality measure. The algorithm's performance was tested on both simulated and actual medical data sets. In both simulation studies and practical medical image segmentation, the system shows promising results in comparison with two well-known methods: the competitive Hopfield neural network (CHNN) and ISODATA methods.; A novel approach for fast registration of two sets of 3D curves or surfaces is also presented. The technique is an extension of Besl and Mackay's iterative closest point (ICP) algorithm. This technique solves the computation complexity associated with the ICP algorithm by applying a novel grid closest point (GCP) transform and a genetic algorithm to minimize the cost function. The GCP transform essentially converts the 3D space surrounding the 2 data sets into a field in which every point stores the magnitude and direction of a displacement vector from this point to the nearest surface element. Thus the cost function is largely precomputed. A detailed description of the algorithm is presented together with a comparison of its performance versus several registration techniques. The algorithm is used to register 2D head contours extracted from CT/MRI data to correct for mis-alignment caused by motion artifact during scanning. Registration using the GCP/GA technique is found to be significantly faster and of comparable accuracy than other techniques that have been developed so far.; As an application, the segmentation and registration algorithms will be combined together in a system that will be applied to extract specific brain structures from traumatic injury patients, namely the ventricles, the corpus callosum, and the pons. These volumes are tracked over time to study their effect in recovery.
展开▼
