Interactive High-Quality MIP

For exact depiction of even small features using MIP, there is need for algorithms which produce high-quality MIP in real-time [41]. In contrast to interactive MIP techniques, which perform just zero-order (nearest-neighbor) interpolation, more accurate evaluation of ray maxima is required for the generation of high-quality MIP (trilinear interpolation, or higher order methods) . At the cost of significantly longer computation times, this allows to create much more detailed images and accurate animations (figure 4.11).

Figure 4.11: MIP with a) shear/warp projection with nearest-neighbor interpolation b) ray casting and trilinear interpolation.

$\includegraphics[width=.48\linewidth]{Figures/swhd.ps}$	$\includegraphics[width=.48\linewidth]{Figures/ihd.ps}$
(a)	(b)

Two major limitations to the performance of current software-based high-quality MIP can be identified. First, although regions of the volume which do not contain meaningful data can be identified in advance and skipped during rendering by the use of distance volumes or similar structures [67], the overhead associated with evaluating the additional information and stepping over these cells significantly limits the possible speed-up of volume traversal. Secondly, despite of space-leaping approaches, the number of interpolations which are actually performed is still far from the optimum. As the volume is traversed in a spatially ordered manner along viewing rays, local maxima are usually encountered and evaluated before the global ray maximum is reached. Moreover, lots of unnecessary evaluations are performed on the rising slopes of data values which precede a maximum.

Based on the principles from chapter 3, a new algorithm for the generation of high-quality MIP (parallel projection) is presented, which is approximately one to two orders of magnitude faster than other software-based approaches with comparable quality. In contrast to the previously described voxel-based MIP, the approach works on cells, which allows to achieve a high quality of the image by performing trilinear interpolation during projection. The projection is done using a ray casting approach. In Section 4.2.1 the preprocessing scheme is presented, which can (but not necessarily has to) be applied to identify and exclude non-contributing cells from the volume. To maximize the amount of cells which do not contribute to any MIP, 12 sets of cells are generated, corresponding to 12 clusters of viewing directions. Usually, more than two thirds of all cells can be eliminated from the data, no longer causing any overhead to identify them later and skip over them. This is achieved by resorting the cells according to their maximum value (Section 4.2.2). As the spatial order of processing cells is not relevant for maximum evaluation, cells containing high data values are evaluated first. This reduces the number of evaluations required for MIP significantly. To avoid even more of the relatively expensive trilinear interpolations, a fast method for estimating the maximum value along a ray-cell intersection is used (section 4.2.3). Using these techniques, the number of trilinear interpolations required per image pixel is greatly reduced, achieving interactive high-quality MIP.