Initial study of anisotropic textures for identification of blood vessels in 7T MRI brain phase images

Within medical science, pattern recognition is the basis for computer-aided diagnosis (CAD), which assists doctors (in particular radiologists) in the interpretation of medical images, whose quality and usefulness are constantly evolving. Modern magnetic resonance imaging (MRI) scans can provide information about subvoxel anatomical structures (e.g. microvessels) that may not be specifically resolvable within a given image set. Hence, subvoxel anatomical structures within a given image may not be readily apparent to the observer (i.e. radiologist); nevertheless, their presence may be detected through their statistical relationship with their surrounding voxels. Such statistical relationships can be characterized by texture features.;The goal of this research dissertation is to investigate the feasibility of using anisotropic texture features for the identification of blood vessels that may not be specifically resolvable in the image datasets. Such features can be used as the basis for the feature extraction component of a complete pattern recognition system for the purpose of automatically identifying blood vessels in the human brain. The approach of this project is to apply 2D statistical texture features as inputs to a classifier (such as a neural network) to analyze MRI images. The specific aims of this dissertation are: (a) to provide a set of texture features extracted from 7T MRI human brain phase images that demonstrate the ability to characterize the presence of underlying microvessel structures; (b) to provide a classifier, in particular a neural network architecture that makes use of the extracted features; and (c) to evaluate the performance of the feature-classifier combination.;The results of this research demonstrated the feature-classifier combination exhibit reasonably well generalization across the testing data, and suggest it may be possible for a computer to discriminate hidden vessels not detectable by human observers.