High-performance image registration algorithms for multi-core processors

Deformable registration consists of aligning two or more 3D images into a common coordinate frame. Fusing multiple images in this fashion quantifies changes in organ shape, size, and position as described by the image set, thus providing physicians with a more complete understanding of patient anatomy and function. In the field of image-guided surgery, for example, neurosurgeons can track localized deformations within the brain during surgical procedures, thereby reducing the amount of unresected tumor. Though deformable registration has the potential to improve the geometric precision for a variety of medical procedures, most modern algorithms are time consuming and, therefore, go unused for routine clinical procedures. This thesis develops highly data-parallel registration algorithms suitable for use on modern multi-core architectures, including graphics processing units (GPUs). Specific contributions include the following:Parallel versions of both unimodal and multi-modal B-spline registration algorithms where the deformation is described in terms of uniform cubic B-spline coefficients. The unimodal case involves aligning images obtained using the same imaging technique whereas multi-modal registration aligns images obtained via differing imaging techniques by employing the concept of statistical mutual information. Multi-core versions of an analytical regularization method that imposes smoothness constraints on the deformation derived by both unimodal and multi-modal registration. The proposed method operates entirely on the B-spline coefficients which parameterize the deformation and, therefore, exhibits superior performance, in terms of execution-time overhead, over numerical methods that use central differencing. The above contributions have been implemented as part of the high-performance medical image registration software package Plastimatch, which can be downloaded under an open source license from www.plastimatch.org. Plastimatch significantly reduces the execution time incurred by B-spline based registration algorithms: compared to highly optimized sequential implementations on the CPU, we achieve a speedup of approximately 21 times for GPU-based multi-modal deformable registration while maintaining near-identical registration quality and a speedup of approximately 600 times for multi-core CPU-based regularization. It is hoped that these improvements in processing speed will allow deformable registration to be routinely used in time-sensitive procedures such as image-guided surgery and image-guided radiotherapy which require low latency from imaging to analysis.