Improving the 3D Visualization of the Visible Korean Human via Data Driven 3D Segmentation in RGB Color Space

The Visible Korean Human (VKH) dataset opens up new possibilities and challenges for computer supported visualization and inspection of three-dimensional anatomical structures. High quality three-dimensional visualizations can be generated using voxel based surface rendering techniques with sub-voxel resolution as available in the VOXEL-MAN System. However, before three-dimensional visualizations of anatomical structures can be generated in high quality an accurate segmentation of the structures has to be performed. Because an automatic segmentation of most anatomical structures is impossible, an interactive edgeoriented segmentation was done for selected structures using the live-wire method. However, small segmentation errors were unavoidable using this interactive slice-oriented technique that induced a significant reduction of the quality of three-dimensional visualizations. In this paper, a method for a data-driven correction of the manual segmentation results is presented and improved visualizations of the skull of the VKH are shown. The method uses knowledge about typical color values occurring in the segmented structure. With the help of three-dimensional morphological operators voxels in the threedimensional neighborhood of the pre-segmented object are selected. These voxel’s (R,G,B) vectors are tested for their similarity to the mean vector of the segmented structure. As similarity measure the Mahalanobis distance is used implicitly respecting the correlation between the R,G,B features. From a geometrical point of view the use of the Mahalanobis distance leads to a characterization of the segmented tissue by an ellipsoid in the (R,G,B) color space. First results show that based on the refined segmentation improved three-dimensional visualizations using voxel based surface rendering of the skull can be generated. The method described has been integrated into the VOXEL-MAN System.