AUTHOR=Shamonin Denis P., Bron Esther E., Lelieveldt Boudewijn P., Smits Marion , Klein Stefan , Staring Marius 

TITLE=Fast Parallel Image Registration on CPU and GPU for Diagnostic Classification of Alzheimer's Disease

JOURNAL=Frontiers in Neuroinformatics

VOLUME=Volume 7 - 2013

YEAR=2014

URL=https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2013.00050

DOI=10.3389/fninf.2013.00050

ISSN=1662-5196

ABSTRACT=Nonrigid image registration is an important, but time-consuming task<br/>in medical image analysis. In typical neuroimaging studies, multiple<br/>image registrations are performed, i.e. for atlas-based segmentation<br/>or template construction. Faster image registration routines would<br/>therefore be beneficial.<br/><br/>In this paper we explore acceleration of the image registration<br/>package elastix by a combination of several techniques: i)<br/>parallelization on the CPU, to speed up the cost function derivative<br/>calculation; ii) parallelization on the GPU building on and<br/>extending the OpenCL framework from ITKv4, to speed up the Gaussian<br/>pyramid computation and the image resampling step; iii) exploitation<br/>of certain properties of the B-spline transformation model; iv)<br/>further software optimizations.<br/><br/>The accelerated registration tool is employed in a study on<br/>diagnostic classification of Alzheimer's disease and cognitively<br/>normal controls based on T1-weighted MRI. We selected 299<br/>participants from the publicly available Alzheimer's Disease<br/>Neuroimaging Initiative database. Classification is performed with a<br/>support vector machine based on gray matter volumes as a marker for<br/>atrophy. We evaluated two types of strategies (voxel-wise and<br/>region-wise) that heavily rely on nonrigid image registration.<br/><br/>Parallelization and optimization resulted in an acceleration factor<br/>of 4-5x on an 8-core machine. Using OpenCL a speedup factor of ~2<br/>was realized for computation of the Gaussian pyramids, and 15-60 for<br/>the resampling step, for larger images. The voxel-wise and the<br/>region-wise classification methods had an area under the<br/>receiver operator characteristic curve of 88% and 90%,<br/>respectively, both for standard and accelerated registration.<br/><br/>We conclude that the image registration package elastix was<br/>substantially accelerated, with nearly identical results to the<br/>non-optimized version. The new functionality will become available<br/>in the next release of elastix as open source under the BSD license.