Simulationsstudien zur Untersuchung der Bildqualität für die 3D-Visualisierung tomografischer Volumendaten

In recent years, three-dimensional (3D) visualization of tomographic volume data has become an important tool for surgical planning and simulation, as well as for building virtual body models and many other applications in the medical field. Results depend on a large number of parameters, often interactively chosen, for image acquisition (e. g. voxel size, slice distance), segmentation (threshold value, transfer functions), and visualization (e. g. interpolation, gradient method). Many of the resulting 3D images look very realistic, due to the methods of computer graphics applied. However, it remains mostly unclear how accurate they really are. This work investigates the image quality of 3D visualization of surfaces from tomographic volume data. First, the very complex and multi-faceted state of the art is reviewed and structured according to the basic aspects of image quality, intelligibility and fidelity. Investigation of image quality obviously requires some knowledge of the true anatomical situation, which is not known for in vivo cases. Therefore, I develop a method based on the simulated tomographic acquisition of geometrically described test scenes. The simulated acquisition is done with due consideration of the partial volume effect. Objects are assumed as being homogeneous, with additive noise. Three test objects are developed, representing various anatomical situations (e. g. thin bones, fractures). Furthermore, two error measures, the localization error and the surface normal error, are defined. They describe the accuracy of the rendered surface. Using these methods, I then vary the major parameters and visualize the local deviations on 3D images and corresponding error images. It is shown that under certain conditions, the localization error is at the order of one tenth of the scanner resolution. These conditions are (1) a diameter of the object to be visualized well above the resolution of the scanner; (2) application of the 50% threshold value; (3) a good contrast-to-noise ratio; and (4) processing according to the sampling theorem. Visualization of small structures is limited by the resolution of the scanner. This also applies to volume rendering. For a selection of processing parameters, I am providing a best practice guide. Determination of a suitable threshold value and thus the reproducibility of the results of 3D visualization remains a major problem. The mean contour gradient which was tested here delivers reproducible, but not necessarily optimal values.