Biomedical image segmentation using multiscale orientation fields

An algorithm for labeling image regions based on pixel-level statistical pattern recognition is presented. The structure of multiscale regions about each pixel is measured by means of isotropic Gaussian filters and by a multiscale orientation field. A redundant feature space representing several aspects of image structure across scale, orientation, and space is created. The segmentation algorithm decides membership of pixels in regions by means of simple statistical pattern recognition methods, such as distance measurement and thresholding. Feature vectors are examined locally to determine region membership; the features incorporate multiscale image structure information. Results of multiscale image segmentations on biomedical images are presented