Image registration by a regularized gradient flow. A streaming implementation in DX9 graphics hardware

摘要

The presented image registration method uses a regularized gradient flow to correlate the intensities in two images. Thereby, an energy functional is successively minimized by descending along its regularized gradient. The gradient flow formulation makes use of a robust multi-scale regularization, an efficient multi-grid solver and an effective time-step control. The data processing is arranged in streams and mapped onto the functionality of a stream processor. This arrangement automatically exploits the high data parallelism of the problem, and local data access helps to maximize throughput and hide memory latency. Although dedicated stream processors exist, we use a DX9 compatible graphics card as a stream architecture because of its ideal price-performance ratio. The new floating point number formats guarantee a sufficient accuracy of the algorithm and eliminate previously present concerns about the use of graphics hardware for medical computing. Therefore, the implementation achieves reliable results at very high performance, registering two 257(2)supercript stop images in approximately 3sec, such that it could be used as an interactive tool in medical image analysis.