Research Group of Prof. Dr. M. Griebel
Institute for Numerical Simulation

Medical Imaging


The interactive visualization of large 3D medical data sets such as CT or MRI scans is still a challenge. The problems mainly arise from the sheer size of the data sets which have to be stored and processed somehow. Volume visualization techniques, such as isosurface extraction, quickly encounter the limits of current graphics hardware.

In order to be able to visualize (and process) these large data sets efficiently, multiresolution techniques and parallelization are necessary. However, usually these two paradigms don't match very well. Another problem in the parallel context is load balancing, that is the problem of processors getting idle at any time during visualization. We developed a parallel multiresolution algorithm with dynamic load balancing which achieves an almost optimal speedup on shared memory parallel computers.


If below pictures appear scrambled, please increase the width of your web browser. Click on the pictures to see an enlarged version.

Above you see adaptive slices along various directions through a 3D MRI scan of a human head (courtesy of MeVis, Bremen). Note that the slicing planes can of course be oriented in any direction. The upper row contains the grids corresponding to the grey scale images in the lower row. The adaptive method requires much less triangles to represent the data with only small error in comparison to a full triangulation (which would be a completely black grid). The multiresolution algorithm allows an online adjustment of the error threshold leading to finer or coarser images with higher or lower triangle counts, respectively.

In these images you see isosurfaces of the same data set. Colors indicate portions of the isosurface that have been extracted a the different processor (3 in the upper row, 6 in the lower row). Note that the portions can change between the images since the load balancing is dynamic and the workload of the processors can change with time depending on the scheduler of the operating system. For the rightmost picture of the upper row the front of the head has been clipped away in order to show its interior.


Above you see transparent isosurface renderings of a tooth, a sheep's heart and a knee (data sets by B. Lorensen from the transfer function bake-off). The isovalues of these images have been selected automatically based on average isosurface normals.


Related Projects

Thomas Gerstner