Fast parallel image registration on CPU and GPU for diagnostic classification of Alzheimer's disease.
Shamonin DP, Bron EE, Lelieveldt BP, Smits M, Klein S, Staring M, Alzheimer's Disease Neuroimaging Initiative. · 2013
Read Fast parallel image registration on CPU and GPU for diagnostic classification of Alzheimer's disease. by Shamonin DP, Bron EE, Lelieveldt BP, Smits M, Klein S, Staring M, Alzheimer's Disease Neuroimaging Initiative. (2013) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc3893567.txt.
Introduction
Denis P. Shamonin, Esther E. Bron, Boudewijn P. F. Lelieveldt, Marion Smits, Stefan Klein, and Marius Staring speed up brain image alignment in Frontiers in Neuroinformatics in 2013. Aligning brain scans to a common frame is vital for measuring shrinkage in Alzheimer's disease, yet the math runs slowly. Their work speeds a known alignment package.
They split the work across many CPU cores and onto the graphics chip, and they exploit tricks in the bend model the package uses. The faster tool then powered a study that sorted Alzheimer's patients from healthy controls by gray matter volume from MRI scans. Two sorting styles, by tiny cube and by region, both leaned on heavy alignment work. The gain shows how raw speed can widen what imaging studies can attempt.
Catalog record
| CREATOR | Shamonin DP, Bron EE, Lelieveldt BP, Smits M, Klein S, Staring M, Alzheimer's Disease Neuroimaging Initiative. |
|---|---|
| YEAR | 2013 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Fast parallel image registration on CPU and GPU for diagnostic classification of Alzheimer's disease By Denis P. Shamonin, Esther E. Bron, Boudewijn P. F. Lelieveldt, Marion Smits, Stefan Klein, Marius Staring Published in: Frontiers in Neuroinformatics DOI: 10.3389/fninf.2013.00050 ABSTRACT Nonrigid image registration is an important, but time-consuming task in medical image analysis. In typical neuroimaging studies, multiple image registrations are performed, i.e., for atlas-based segmentation or template construction. Faster image registration routines would therefore be beneficial. In this paper we explore acceleration of the image registration package elastix by a combination of several techniques: (i) parallelization on the CPU, to speed up the cost function derivative calculation; (ii) parallelization on the GPU building on and extending the OpenCL framework from ITKv4, to speed up the Gaussian pyramid computation and the image resampling step; (iii) exploitation of certain properties of the B-spline transformation model; (iv) further software optimizations. The accelerated registration tool is employed in a study on diagnostic classification of Alzheimer's disease and cognitively normal controls based on T1-weighted MRI. We selected 299 participants from the publicly available Alzheimer's Disease Neuroimaging Initiative database. Classification is performed with a support vector machine based on gray matter volumes as a marker for atrophy. We evaluated two types of strategies (voxel-wise and region-wise) that heavily rely on nonrigid image registration. Parallelization and optimization resulted in an acceleration factor of 4–5x on an 8-core machine. Using OpenCL a speedup factor of 2 was realized for computation of the Gaussian pyramids, and 15–60 for the resampling step, for larger images. The voxel-wise and the region-wise classification methods had an area under the receiver operator characteristic curve of 88 and 90%, respectively, both for standard and accelerated registration. We conclude that the image registration package elastix was substantially accelerated, with nearly identical results to the non-optimized version. The new functionality will become available in the next release of elastix as open source under the BSD license. 1. INTRODUCTION Image registration is a frequently used technique in medical image processing. It refers to the process of automatically aligning imaging data, where a moving (target) image IM is deformed to mimick a fixed (reference) image IF. In other words, registration is the problem of finding a coordinate transformation T that makes IM(T) spatially aligned with IF. The quality of alignment is defined by a cost function . The optimal coordinate transformation is estimated by minimizing the cost function with respect to T, usually by means of an iterative optimization method embedded in a hierarchical (multiresolution) scheme. Extensive reviews on the subject of image registration are given in Brown (1992); Maintz and Viergever (1998). Areas of application include the alignment of data sets from different modalities (Mattes et al., 2003) to fuse information, comparison of follow-up with baseline scans (Staring et al., 2007) to follow disease development, alignment of different MR sequences for extraction of quantitative MR parameters such as in diffusion tensor imaging or MR relaxometry (Alexander et al., 2001; Bron et al., 2013), alignment of pre- and post-contrast images (Rueckert et al., 1999) to aid breast cancer detection and diagnosis, and updating treatment plans for radiotherapy and surgery (Pennec et al., 2003). Accordingly, most neuroimaging research also requires image registration. Registration is mainly needed to create a reference frame, which enables comparison between subjects, between image sequences and over time. This reference framework can either be a common template space to which every subject's image is registered (Mazziotta et al., 1995; Seghers et al., 2004; Ashburner, 2007), or a region-labeling system for example obtained with multi-atlas segmentation (Heckemann et al., 2006). Many different neuroimaging applications rely on such a reference framework: statistical group comparisons (Friston et al., 1994), voxel-based morphometry (Ashburner and Friston, 2000), tissue segmentation (Fischl et al., 2002; Ashburner and Friston, 2005), and diagnostic classification (Klöppel et al., 2008; Magnin et al., 2009; Cuingnet et al., 2011). In these applications, registration methods are used to align the data with the reference frame. To create a reference frame that maps between different subjects, nonrigid image registration is applied, which can be very time-consuming. Runtime depends on the specific cost function, transformation complexity, data size, and optimization strategy. The first three items have increased in complexity over the years: more complex cost functions were needed for multi-modal image registration (Maes et al., 1997), nonrigid transformations have many parameters frequently generating a 106 dimensional space to be optimized, and data sizes have increased tremendously with the advent of new scanners. This results in a typical runtime of registration algorithms in the order of at best 15 min, up to hours (Klein et al., 2009a); future acquisition-side improvements in image resolution may even increase that number. Moreover, for creating a reference frame, many registrations are required: every subject needs to be aligned with the template space, or, when using multi-atlas segmentation, every atlas image needs to be aligned with every subject image. One of the neuroimaging applications mentioned above is diagnostic classification. As the incidence of Alzheimer's Disease (AD) as well as the need for early and accurate diagnosis is dramatically growing (Alzheimer's Association, 2012), automated classification is an emerging research field. To advance the diagnosis of AD in individual patients, machine-learning techniques can be applied to imaging or other data. These techniques use labeled data to train a classifier to categorize two groups (e.g., patients and controls). Several studies demonstrated the successful classification of dementia based on atrophy using such machine-learning methods (e.g., Fan et al., 2008; Klöppel et al., 2008; Vemuri et al., 2008; Magnin et al., 2009; Cuingnet et al., 2011; Koikkalainen et al., 2012). The atrophy features used in these studies are derived from structural MR using two main approaches: voxel-wise (e.g., Klöppel et al., 2008) and region-wise (e.g., Magnin et al., 2009) feature extraction. Voxel-wise methods use a feature for each voxel in the brain, for example the gray matter (GM) density as an atrophy measure. In the region-wise approach, a region-labeling consisting of a set of brain regions is used to calculate a feature, for example the GM volume in each region of interest (ROI). Both approaches require many nonrigid image registrations: in the voxel-wise approach, to align all scans in a template space, and in the region-wise approach, to obtain a region-labeling for each individual scan using multi-atlas segmentation. In this paper we explore the acceleration of image registration in the context of neuroimaging applications, by a combination of methods. Critical registration components are parallelized, utilizing the CPU as well as the GPU, certain properties of the B-spline transformation model are exploited, and source code is optimized. These efforts are integrated in the popular open source registration toolkit elastix (Klein et al., 2010), which is based on the Insight ToolKit (ITK, (Ibánez et al., 2005)). elastix aims to deliver convenient access to a wide range of image registration algorithms to end-users (researchers as well as medical practitioners). For the GPU implementation, the recently introduced OpenCL functionality in ITKv4 was improved, extended and exploited. The developed functionality will become available in the next release of elastix, as open source under the BSD license. Others have also addressed registration performance by means of parallel processing. An overview of both CPU and GPU work is