In this paper, we propose a computationally efficient method for medical image registration. The centerpiece of the approach is to reduce the dimensions of each image via a projection operation. The two sequences of projection images corresponding to each image are used for estimating the registration parameters. Depending upon how the projection geometry is setup, the lower dimension registration problem can be parameterized and solved for a subset of parameters from the original problem. Computation of similarity metrics on the lower dimension projection images is significantly less complex than on the original volumetric images. Furthermore, depending on the type of projection operator used, one can achieve a better signal to noise ratio for the projection images than the original images. In order to further accelerate the process, we use Graphic Processing Units (GPUs) for generating projections of the volumetric data. We also perform the similarity computation on the graphics board, using a GPU with a programmable rendering pipeline. By doing that, we avoid transferring a large amount of data from graphics memory to system memory for computation. Furthermore, the performance of the more complex algorithms exploiting the graphics processor’s capabilities is greatly improved. We evaluate the performance and the speed of the proposed projection based registration approach using various similarity measures and benchmark them against an SSE-accelerated CPU based implementation.     «

In this paper, we propose a computationally efficient method for medical image registration. The centerpiece of the approach is to reduce the dimensions of each image via a projection operation. The two sequences of projection images corresponding to each image are used for estimating the registration parameters. Depending upon how the projection geometry is setup, the lower dimension registration problem can be parameterized and solved for a subset of parameters from the original problem. Compu...     »