Volume Segmentation , i.e. the identification of individual objects in volumetric datasets has been an area of extensive research, especially in medical imaging applications. Several algorithms have been proposed and examined with varying qualities of results. Most of the algorithms require a lot of computations. As they work on three-dimensional, large datasets, the execution is slow in the majority of cases and furthermore often requires several passes.

Recent advances in computational power and flexible programming of graphics hardware have expanded their application area to general purpose computations. In addition to high-quality scientific visualization, also segmentation algorithms were implemented on the GPU during the last year. This usage of the GPU (graphics processing unit) leads to interactive segmentation where the progress of the segmentation can be tracked and influenced interactively.

During the last decade nonlinear diffusion models have shown to be powerful tools in image processing, particularly for image smoothing and enhancing, mostly applied to two-dimensional images. A new application that uses diffusion filtering for volume segmentation purposes was presented only recently. The fundamental idea behind this approach is an expansion of a small starting region of interest where the growth is governed by the diffusivity of the underlying data. The algorithm was implemented on graphics hardware leveraging its parallel processing capabilities.

The segmentation tool was integreted into an existing, complex hardware based volume renderer. We describe how this segmentation approach can leverage the parallel processing capabilities of modern graphics hardware and what kind of optimization can be performed during computation in order to obtain interactivity. Different diffusion models like the isotropic- and anisotropic diffusion are presented and different nonlinear diffusivities are investigated, analysing their effects on three-dimensional image smoothing and segmentation. An extension to the isotropic diffusion equation for segmentation is presented which provides similar or even better results at faster rates.

3D Image Smoothing. Nonlinear Isotropic Diffusion after 4, 10 and 25 iterations compared with the original DVR rendered data volume (left).

The same data volume but isosurface rendered. Anisotropic nonlinear diffusion (top row) is applied on the original data (top left). For smooth surfaces, the isotropic nonlinear diffusion (bottom row) has more impact than the anisotropic case.

For noisy data, the anisotropic diffusion removes the noise and preserves the structure. Whereas with the isotropic diffusion the structure gets damaged after a few iterations.

Uniform white noise is applied on the original dataset and isosurface rendered with GPU raycasting (top left image). Anisotropic diffusion is applied and shown after 8, 14 and 24 iterations (top row) and isotropic diffusion after 2, 5, 7 and 15 iterations.

Segmentation

Segmentation of the air in a colon download video.

Myocardium of the left ventricle The upper left image shows the gradient magnitudes of one slice of a CT dataset of the myocardium of the left ventricle. After smoothing with a nonlinear diffusion filter the gradient magnitudes clearly show the borders of the individual objects (bottom left), so that a segmentation with a nonlinear diffusion process is possible (3D view and slice view on the right).

Segmentation of the bone, the paranasal sinuses and the frontal sinus of a head dataset. Different transfer functions are applied on the different segmented object masks. The bone is segmented with simple thresholding and the paranasal- and frontal sinuses with the diffusion-based seeded region growing technique.