Classification of Material Mixtures in Volume Data for Visualization and Modeling

David H. Laidlaw, Kurt W. Fleischer, Alan H. Barr
California Institute of Technology Computer Science Department
Technical Report number CS-TR-94-07


Material classification is a key step in creating computer graphics models and images from volume data. We present a new algorithm for identifying the distribution of different material types in volumetric datasets such as those produced with Magnetic Resonance Imaging (MRI) or Computed Tomography (CT). The algorithm assumes that voxels can contain more than one material, e.g. both muscle and fat; we wish to compute the relative proportion of each material in the voxels.

Other classification methods have utilized Gaussian probability density functions to model the distribution of values within a dataset. These Gaussian basis functions work well for voxels with unmixed materials, but do not work well where the materials are mixed together. We extend this approach by deriving non-Gaussian "mixture" basis functions.

We treat a voxel as a volume, not as a single point. We use the distribution of values within each voxel-sized volume to identify materials within the voxel using a probabilistic approach.

The technique reduces the classification artifacts that occur along boundaries between materials. The technique is useful for making higher quality geometric models and renderings from volume data, and has the potential to make more accurate volume measurements. It also classifies noisy, low-resolution data well.