High Quality Rendering of Attributed Volume Data Ulf Tiede Thomas Schiemann Karl Heinz H ¨ ohne Institute of Mathematics and Computer Science in Medicine University Hospital Eppendorf, Hamburg Abstract For high quality rendering of objects segmented from tomographic volume data the precise location of the boundaries of adjacent ob- jects in subvoxel resolution is required. We describe a new method that determines the membership of a given sample point to an object by reclassifying the sample point using interpolation of the original intensity values and searching for the best fitting object in the neigh- bourhood. Using a ray-casting approach we then compute the sur- face location between successive sample points along the viewing- ray by interpolation or bisection. The accurate calculation of the object boundary enables a much more precise computation of the gray-level-gradient yielding the surface normal. Our new approach significantly improves the quality of reconstructed and shaded sur- faces and reduces aliasing artifacts for animations and magnified views. We illustrate the results on different cases including the Visible-Human-Data, where we achieve nearly photo-realistic im- ages. CR Categories: I.3.3 [Computer Graphics]: Picture/Image Generation—Display algorithms; I.3.7 [Computer Graphics]: Three-Dimensional Graphics and Realism—Color, shading, shad- owing and texture Keywords: partial-volume-effect, ray-casting, tomographic data, Visible-Human-Project 1 INTRODUCTION Three-dimensional (3D) visualization of tomographic volume data such as Computer-Tomography (CT) or Magnetic-Resonance- Imaging (MR) is an important aid for diagnosis, treatment plan- ning, teaching and other applications. The common feature of these imaging modalities is that different objects are characterized as different levels of the measured intensity values. At the bound- ary of adjacent objects the intensity jumps from one level to the other. However, due to the limited spatial resolution of the scan- ning devices the intensity of the voxels containing the boundary is averaged. This effect is called the partial-volume-effect. A variety of 3D-visualization methods using this effect implicitly have been published, e.g. isosurface reconstruction and http://www.uke.uni-hamburg.de/idv volume-rendering One well-known algorithm for isosurface reconstruction is Marching-Cubes [7] which produces polygonal meshes of the struc- tures that are within a given intensity range. The polygonal object representation can be rendered very fast on special graphics hard- ware. The disadvantage is that only surfaces can be displayed, be- cause the information about the interior of the object contained in the original intensity values is lost. However, these values might be of great concern for diagnostic purposes. In the volume-rendering approach originally investigated by [1, 6] an opacity value is computed for each voxel by assigning color or transparency values to the original gray values and weighting them with the magnitude of the gray-level-gradient at the voxel position. A voxel with a high opacity is less transparent than a voxel with a lower value. During ray-casting the voxels are shaded and weighted with their respective opacity yielding multiple transparent surfaces, which are difficult to interpret [13]. Although volume-rendering produces nice looking images most medical applications require an explicit object representation that allows measurements and removal of structures which obscure the view to the objects of interest. In many cases objects can be differentiated with distinctive inten- sity ranges, a method known as threshold segmentation. Typically the average intensity value of adjacent objects is used for thresh- olding. An intensity range for an object is then specicified using a lower and an upper threshold value. However, if different objects share the same intensity range an explicit assignment which voxel belongs to which object is required. We use morphological oper- ations and connected-component-analysis as described in [4]. In difficult cases the segmentation has to be done manually. The re- sult of the segmentation process is an additional volume containing discrete attribute values for each voxel about its membership to a certain organ. These attribute values trigger the individual visual- ization of the respective objects. Problem Ray-casting for visualizing sampled volume data has been elabo- rately described in the literature [5, 11, 15]. The classical approach uses discrete scan-conversion algorithms to access the attribute val- ues which control the visualization of the objects. The pitfall of this method is that the subvoxel resolution of the viewing ray is lost which causes staircase looking surfaces of the objects (fig. 1 left). For technical reasons or patient care the enhancement of the spatial resolution of the volume data to reduce this effect is not possible in most cases. Although the real boundary of an object cannot be reconstructed exactly, for realistic and high resolution renderings of attributed volume data a better approximation of the surface than the arbitrary voxel boundaries is required (fig. 1 right). None of the methods published so far accomplish this for segmented data.