GPU Accelerated Normalized Mutual Information and B-Spline Transformation

Abstract

Visualization of multimodal images in medicine and other application areas requires correct and efficient registration. Optimally, the alignment operation is made an integral part of the rendering process. Voxel based approaches using mutual information ensure high quality similarity measurement. As a limiting factor, high computational load is caused since for every iteration of the optimization procedure one volume is transformed into the coordinate system of the other, a 2D histogram is generated and mutual information is computed. The expensive trilinear interpolation operations are well supported by 3D texture mapping hardware. However, existing strategies compute the histogram and mutual information on the CPU and thus require a cost intensive data transfer. Overcoming this considerable bottleneck, we introduce a new approach that efficiently supports all computations on modern graphics cards. This makes expensive data transfers from GPU to main memory dispensable. Due to its modularity, the presented approach can be integrated into general frameworks. As a major result, the speed improvement over existing GPU-CPU strategies amounts to a factor of 4 and over pure conventional CPU techniques to more than a factor of 10. Overall, the suggested strategy contributes considerably to integration of multimodal registration directly into interactive volume visualization.