Report Number: CS-TR-99-1617
Institution: Stanford University, Department of Computer Science
Title: Segmentation of Medical Image Volumes Using Intrinsic Shape Information
Author: Shiffman, Smadar
Date: February 1999
Abstract: I propose a novel approach to segmentation of image volumes that requires only a small amount of user intervention and that does not rely on prior global shape models. The approach, intrinsic shape for volume segmentation (IVSeg), comprises two methods. T he first method analyzes isolabel-contour maps to identify salient regions that correspond to major objects. The method detects transitions from within objects into the background by matching isolabel contours that form along the boundaries of objects as a result of multilevel thresholding with a fine partition of the intensity range. The second method searches in the entire sequence for regions that belong to an object that the user selects from one or a few sections. The method uses local overlap criter ia to determine whether regions that overlap in a given direction (coronal, sagittal, or axial) belong to the same object. For extraction of blood vessels, the method derives the criteria dynamically by fitting cylinders to regions in consecutive sections and computing the expected overlap of slices of these cylinders. In a formal evaluation study with CTA data, I showed that IVSeg reduced user editing time by a factor of 5 without affecting the results in any significant way.