GPU-Accelerated Tools for Medical Image Registration and Biomechanical Modeling

Abstract

This dissertation explores the implications of GPU-accelerated computing (i.e., harnessing the parallel computing power of modern graphics processing units) for applications in image registration and biomechanical modeling. The work described herein includes the development of two software toolkits: one that implements functions for anatomical modeling based on distance fields, and one for registration of 3D (volumetric) medical imaging with 2D imaging (including stereo pairs). In both cases, the general aim of the research effort was to provide practical support for creating and simulating anatomical models with greater accuracy and efficiency; the specific goals involved applications of interest of a research group focused on understanding the biomechanics of the human foot. Thus, the distance-based tools were employed for purposes such as enhancing an existing finite-element model of the foot by identifying appropriate locations for cartilage elements that cannot be reliably resolved from typical imaging studies. The primary application described for the 2D-3D registration toolkit involves high-accuracy markerless tracking of the kinematics of the bones in the foot during walking gait. In both cases, GPU-based parallelism was found to have a significant impact on software execution time for data-intense applications and succeeded in realizing gains in computational efficiency of 10- to 150-fold over traditional CPU-based computing. The ramifications of such a radical increase in raw computing power are many, but possibly one of the most important outcomes for the group is the ability to rapidly and accurately quantify bone motion in the human foot during gait. Moreover, this research will deepen our understanding of foot bone motion as it relates to subjects exhibiting both normal and abnormal (i.e., deformities) foot conditions. Previous attempts at engaging in studies involving a large subject pool were stifled by prohibitively long software execution times; however, the GPU-based image processing tools developed in this dissertation will enable our group to revisit the once impractical investigation. While the human foot is the anatomical region of choice for studies described in this dissertation, the tools presented can be adapted to other anatomical localities in the body, such as the knee, hip, hand or shoulder. Additionally, this toolbox need not be limited to the human anatomy, opening the door to efficient exploration of animal models. Simulation of anatomically correct models can help investigators to more accurately predict physiological function, diagnose disease, and analyze environmental impact, which may lead to higher success rates during treatment and recovery. The GPU-based tools presented in this dissertation will provide the foundation for future work in the field of computational biomechanics by arming investigators with resources that encompass particularly relevant parallel computing tools. The intent is to provide an intuitive and interactive environment for efficiently modeling and manipulating highly complex biological geometry.