3D Segmentation and Labeling Using Unsupervised Clustering For Volumetric Measurments On Brain CT Imaging

In this paper, we present a new system to segment and label CT Brain slices using diierential geometrical invariant features and unsupervised clustering. Our aim is to extract reliable and robust measures from CT images of Traumatic Brain Injury (TBI) patients that can accurately describe the morphological changes in the brain as recovery progresses, and to study the correlation between the morphological changes in progressive CT images after the Day Of Injury (DOI) with cognitive functional changes. Segmentation is performed by assigning a feature pattern to each voxel, consisting of a scaled family of diierential geometrical invariant features. The invariant feature pattern is assigned to a certain region using unsupervised clustering. In this paper we used 3 clustering techniques: The ISODATA Algorithm, the Hoppeld network, and the Kohonen feature map network. Implementation and performance of these techniques are presented.