A fast surface-aware 3D non-linear image registration algorithm implemented on a GPU

A novel regularization algorithm has been implemented in our previously described image registration program Ezys. The performance of the CPU enabled package is still very fast with typical 3D image registrations taking less than a minute. The method is based on a surface detection in images of interest. It allows more degrees of freedom near detected surfaces to obtain a good match of them, and enforces more smoothness elsewhere. The new regularization algorithm helped to increase the resolution of the resulting Jacobian (volume change) maps significantly and suppress noise in regions with few anatomical landmarks. The algorithm was also tested on longitudinal data of subjects with Alzheimer´s disease, Semantic dementia and a control group with promising results in quantifying longitudinal volume changes in a timescale of 1 year.