Dan Adler's Master's Thesis

Registration of three-dimensional images is an important task in biomedical science. However, the computational costs of 3D registration algorithms have hindered their widespread use in many clinical and research workflows. Mr. Adler’s research focused on developing an automated medical image registration framework that leverages the computational power of commodity graphical processing units (GPUs) to greatly accelerate this task. His methods take advantage of the graphics hardware’s high computational parallelism and memory bandwidth to perform affine, intensity-based registration of multi-modal 3D medical images at near interactive rates. He also developed a GPU implementation of a common deformable 3D registration technique (Demons algorithm). This algorithm is capable of performing complex non-linear warps to align medical volumes from different individuals in 40 seconds or less. He also developed a GPU implementation of a powerful algorithm to measure the quality of, and guide, 3D volume alignment (Normalized Mutual Information).

His GPU algorithms were compared to state-of-the-art serial algorithms in a series of experiments that employed phantom and patient images. In each experiment the co-registration algorithms attempted to recover a known linear or non-linear transformation. The GPU algorithms had accuracy equivalent to state-of-the-art serial algorithms, but computed their results 10x to 20x more quickly on average.

The GPU algorithms were then used in a clinical study to examine the relationship between the genetic makeup and spatial distribution of glioblastoma multiforme brain tumors. These aggressive tumors are highly resistant to treatment. New information about their growth patterns and genetics may help scientists devise new treatments.