3D Volume Segmentation of MRA Data Sets Using Level Sets Aly A. Farag, Ph.D. 1 , Hossam Hassan, M.S. 1 , Robert Falk, M.D. 2 and Stephen G. Hushek, Ph.D. 3 1 Computer Vision and Image Processing Laboratory, University of Louisville, Louisville, KY 40292 2 Director, Medical Imaging Division Jewish Hospital, Louisville, Kentucky 3 Technical Director, IntraOperative MRI Center, Norton Hospital, Louisville, Kentucky Contact Author: Aly A. Farag, CVIP Lab, Rm 412 Lutz Hall, Louisville, KY 40292. E-mail: farag@cvip.uofl.edu ; URL: www.cvip.uofl.edu Abstract. In this paper, we use a level set based segmentation algorithm to extract the vascular tree from Magnetic Resonance Angiography (MRA) data sets. The classification approach depends on initializing the level sets in the 3D volume and, the level sets evolve with time to yield the blood vessels. This work intro- duces a high quality initialization for the level set functions, allowing extraction of the blood vessels in 3D and elimination of non-vessel tissues. A comparison between the 2D and 3D segmentation approaches is made. The results are vali- dated using a phantom that simulates the MRA data and demonstrate good accu- racy. 1. Introduction The human cerebrovascular system is a complex three-dimensional anatomical structure. Se- rious types of vascular diseases such as carotid stenosis, aneurysm, and vascular malforma- tion may lead to brain stroke, which is the third leading cause of death and the number one cause of disability. An accurate model of the vascular system from MRA data volume is needed to detect these diseases at early stages and prevent invasive treatments. A variety of methods have been developed for segmenting vessels within MRA data. One class of meth- ods is based on a statistical model, which classifies voxels within the image volume into ei- ther vascular or non-vascular classes from time-of-flight MRA [1]. Another class of segmen- tation is based on an intensity threshold where points are classified as either greater or less than a given intensity. This is the basis of the iso-intensity surface reconstruction method [2]- [4]. This method suffers from errors due to image inhomogeneities and the choice of the threshold level is subjective. An alternative to segmentation is axis detection known as the skeletonization process, where the central line of the tree vessels is extracted based on the