high performance computing on graphics processing units: hgpu.org

Image registration is a fundamental step in many applications involving image analysis. It consists of optimizing a similarity metric to find a spatial transformation to match two images (in 3D). It has application in medical images to build atlases (registering a population), or to align a patient to a template to detect pathologies. The main objective of this project is to investigate different image registration algorithms in the context of developing a cloud computing service that it will be used in medical images application. In order to reach this goal, at first we ported in C++ simple block matching algorithms, we used a local brute force searching for the block in the fixed image that best matches each block within the moving image. Several different classes of C++ are built for the matching processing only based on rigid transformation. At first we used the sequential block matching algorithm, however its complexity is too expensive, so we used the parallel block matching system using OpenCL library (Graphics Processing Unit programming). The result is a list of matched coordinates which could be seen as a list of local transformations. Fo each local transformation we have a score which shows the trust on that transformation. With these results we can then find the global rigid transformation between both images.