Automated initialization and automated design of border detection criteria in edge-based image segmentation

This thesis provides methodology for fully automated model-based image segmentation. All information necessary to perform image segmentation is automatically derived from a training set that is provided in a form of segmentation examples. The training set is used to construct two models representing the objects—Shape Model and Border Appearance Model.; We propose a two-step approach to image segmentation. In the first step, an approximate location of the object of interest is determined. In the second step, accurate border segmentation is performed. The approximate location found in the first step is used to either create a region of interest or to otherwise initialize the algorithm that performs the accurate image segmentation.; Active Hough Transform methodology was developed that provides accurate initialization automatically. It finds objects of arbitrary shape, rotation or scaling and can handle object variability. A Border Appearance Model was developed to automatically design cost function that can be used in the segmentation criteria of any edge-based segmentation method. The automated design of cost function greatly simplifies and significantly speeds up design of image segmentation criteria for new segmentation applications.; Our method was tested in five different segmentation tasks that included 489 objects to be segmented (endocardial and epicardial borders in MR images of thorax, Corpus Callosum and Cerebellum in MR images of brain and vertebrae in MR images of spine). The automated detection of the approximate object location always succeeded in providing an accurate initialization [rms errors in pixels: 2.4 (Cerebellum), 1.9 (Corpus Callosum), 2.4 (vertebrae), 1.6 (epicardial) and 2.1 (endocardial) borders]. The automatically designed cost function was applied to the Dynamic Programming and to the Snakes segmentation algorithms to demonstrate general applicability of our approach. The segmentation was compared to manually defined borders with good results [rms errors in pixels: 1.2 (Cerebellum), 1.1 (Corpus Callosum), 1.5 (vertebrae), 1.4 (epicardial) and 1.6 (endocardial) borders].; This work solves two major problems of the state-of-the-art edge-based image segmentation algorithms: strong dependency on a close-to-target initialization, and necessity for manual redesign of segmentation criteria whenever new segmentation problem is encountered.