KNOWLEDGE-BASED SEGMENTATION AND STATE-BASED CONTROL IN IMAGE-ANALYSIS - 2 EXAMPLES FROM THE BIOMEDICAL DOMAIN

This paper presents a system for automatic image analysis featuring kn owledge-based segmentation and state-based control. It is assumed that the image depicts known scenes with objects characterized by well-def ined spatial relations and fuzzy boundaries. The image is preprocessed , the results are stored in a multiresolution pyramid, and segmentatio n primitives are stored as feature vectors. Regions are labeled with a simple model of direction and distance relations in the scene. The la beling procedure affects the segmentation: if the labeled regions in t he image are incompatible with the model, the segmentation and labelin g can be reiterated on a higher resolution, or objects can be removed using a maximal inconsistency criterion. A state-based control mechani sm is introduced; a scheduler initiates a set of processing states in an execution evaluation cycle, and a diagnosis procedure evaluates the current results and determines which processing state to pursue at th e next timestep. Two image domains are described, the delineation of t he left ventricle in four-chamber ultrasound cardiac images, and gamma camera images of the heart.