Abstract
The segmentation of images with poor contrast characteristics is an important issue in Medical Computer Vision. Often image segmentation results are either oversegmented, with “objects”" divided into parts, or incorrectly segmented, with two or more anatomies segmented as one single object. This problem occurs in all types of segmentation approaches, but is of particular importance in the field of region-growing algorithms, which are used in many medical applications, preventing the definition of stable and reliable segmentation parameters. We present a new knowledge-based method, based on an extension of the inexact consistent labeling method, that enables the automated consistency checking of the results of region-growing segmentations and that is capable to automatically “fitting” erroneous segmentations, when they are oversegmented, given there exists a reliable domain model that can be used to guide a tree search procedure in the labeling space. This allows the use of oversensitive parameters always when an exact segmentation is not reliable.