| Description |
Stochastic methods, dense free-form mapping, atlas construction, and total variation are examples of advanced image processing techniques which are robust but computationally demanding. These algorithms often require a large amount of computational power as well as massive memory bandwidth. These requirements used to be ful lled only by supercomputers. The development of heterogeneous parallel subsystems and computation-specialized devices such as Graphic Processing Units (GPUs) has brought the requisite power to commodity hardware, opening up opportunities for scientists to experiment and evaluate the in uence of these techniques on their research and practical applications. However, harnessing the processing power from modern hardware is challenging. The di fferences between multicore parallel processing systems and conventional models are signi ficant, often requiring algorithms and data structures to be redesigned signi ficantly for efficiency. It also demands in-depth knowledge about modern hardware architectures to optimize these implementations, sometimes on a per-architecture basis. The goal of this dissertation is to introduce a solution for this problem based on a 3D image processing framework, using high performance APIs at the core level to utilize parallel processing power of the GPUs. The design of the framework facilitates an efficient application development process, which does not require scientists to have extensive knowledge about GPU systems, and encourages them to harness this power to solve their computationally challenging problems. To present the development of this framework, four main problems are described, and the solutions are discussed and evaluated: (1) essential components of a general 3D image processing library: data structures and algorithms, as well as how to implement these building blocks on the GPU architecture for optimal performance; (2) an implementation of unbiased atlas construction algorithms|an illustration of how to solve a highly complex and computationally expensive algorithm using this framework; (3) an extension of the framework to account for geometry descriptors to solve registration challenges with large scale shape changes and high intensity-contrast di fferences; and (4) an out-of-core streaming model, which enables developers to implement multi-image processing techniques on commodity hardware. |
