Fast Surface Rendering for Three Dimensional Medical Image Reconstruction

Surface rendering, together with volume rendering, is the most popular sort of algorithm used in scientific visualization. Though developed later than surface rendering and having a wider adaptability to object data sets, the algorithm of volume rendering still can not be put into use in practical medical imaging application,simply because it could not meet the speed requirement for interactive operation due to the large amount of computation involved.Surface rendering thus becomes the general method for visualization in most medical imaging, for it is at least more computation-economic than its counterpart. The common algorithm of surface rendering can be generalized by two main steps - isosurface extracting and isosurface shading. Of all methods proposed for isosurface extracting, Marching Cubes (or Dividing Cubes) is the most widely used one. To pick up an isosurface for display, it traverses all the cells in the data set, and for each cell it compares the included eight vertexes to a certain threshold to decide whether it is a boundary cell or not. However, extensive investigation shows that, of all the units checked out, there are usually only less than 10% are useful. Considering the inefficiency in the search for boundary cells the whole run time can be further reduced by applying an optimized search order. Because each isosurface we are extracting from the volume data set actually corresponds to the epidermis of a human organ, we assume that the isosurfaces are continual as well. With the continuity assumption, the isosurface extracting can be accomplished by two steps. First, the volume data set are subsampled by 1/8 voxel.After comparing with the threshold, all the voxels in the subset are classified into boundary points and non-boundary points. After the search for isosurface is carried out, but only in the neighboring space of boundary points, those voxels labeled as non-boundary points will be completely neglected for further processing. As a result, the optimized algorithm is at least 4 times more efficient, with the possibility that the rendered image be mildly distorted by an average error of less than 1 voxel in all directions.When applied to practical use - usually large image series, it is foreseeable that the distortion will be so slight as nearly undetectable. Since surface fitting alone can not produce any image usable, the segmentation of volume data has to be discussed. For many CT image series, the separation of different tissues such as bones, teeth from brain could be accomplished conveniently by setting an apposite threshold, but so far as the MRI and B scan image series are concerned, other more elaborate methods should be applied. This paper discusses the improvement of surface rendering for practical use. In section 1, the basic principle and rendering scheme are presented. Section 2 deals with the volume data segmentation. Then in section 3, an optimized isosurface searching order is proposed. Section 4 discusses isosurface shading. In section 5 the rendering speed of the new method is provided with generated image in contrast to the original one.