3D segmentation of Ct images on a massively parallel computer

Abstract

For 3D scenes a 3D segmentation technique on a massively parallel computer is described and tested on CT image sequences. The approach is based on a volume growing technique driven by statistical features and a model depending on characteristic object parameters. The volumes of interest are specified interactively and used as seed volumes for the growing algorithm. An estimation technique is employed to calculate several statistical properties of these seed volumes. The required homogeneity criterion for each volume is then obtained in regard to the estimated statistics and the model of the object. These segmentation results are handled by a 3D morphological operator. Due to practical considerations the approach has been implemented on a massively parallel SIMD (single instruction multiple data) machine, the MasPar Mp1102.