Evaluation of volume-based and surface-based brain image registration methods

Abstract

Establishing correspondences across brains for the purposes of comparison and group analysis is almost universally done by registering images to one another either directly or via a template. However, there are many registration algorithms to choose from. A recent evaluation of fully automated nonlinear deformation methods applied to brain image registration was restricted to volume-based methods. The present study is the first that directly compares some of the most accurate of these volume registration methods with surface registration methods, as well as the first study to compare registrations of whole-head and brain-only (de-skulled) images. We used permutation tests to compare the overlap or Hausdorff distance performance for more than 16,000 registrations between 80 manually labeled brain images. We compared every combination of volume-based and surface-based labels, registration, and evaluation. Our primary findings are the following: 1. de-skulling aids volume registration methods; 2. custom-made optimal average templates improve registration over direct pairwise registration; and 3. resampling volume labels on surfaces or converting surface labels to volumes introduces distortions that preclude a fair comparison between the highest ranking volume and surface registration methods using present resampling methods. From the results of this study, we recommend constructing a custom template from a limited sample drawn from the same or a similar representative population, using the same algorithm used for registering brains to the template.

Introduction

Brain images are registered to one another (or to an average template) to establish correspondences of all kinds, such as across structures, patterns of functional activity, physiological data, and connectivity data. These correspondences enable comparison across time, task, and population. Brain image registrations are performed either on image volumes in their “native” space, or on surface representations of the brain.

Surface registration methods require computationally intensive extraction of a cortical surface, and may not be accurate for topologically different brains (such as lesioned and other pathological cases), but have been demonstrated to perform accurately under many conditions, including recent studies comparing surface features with cytoarchitectonic data (Rademacher et al., 1993, Hinds et al., 2008, Fischl et al., 2008). Significant advantages of performing registrations on a surface compared with in a volume include computational efficiency (less to register, one less degree of freedom), and distances along the cortical surface are more faithfully represented as geodesic distances along a surface rather than Euclidean distances across, for example, banks of a sulcus.

Most prior attempts to compare volume and surface registration methods have used function to gauge registration accuracy and compared a nonlinear surface-based spherical registration method with affine, Talairach-based linear or piecewise linear registration (Anticevic et al., 2008, Desai et al., 2005, Fischl et al., 1999, Thompson and Toga, 1996). Hinds et al. (2008) compared the quality of two atlases, one constructed using FreeSurfer spherical registration and the other using a single volume-based nonlinear registration method, according to the cumulative probability of a region (V1); this atlas comparison is a more indirect comparison than evaluating pairwise registrations.

In this study, we attempt to directly compare surface with volume registration methods. For the volume registration methods, we selected SyN (Avants et al., 2008) and ART (Ardekani et al., 1995, Ardekani et al., 2005), the only methods that attained top rank for all tests and for all four label sets and populations used in a recent, large-scale evaluation of brain image registration methods (Klein et al., 2009).

For the surface registration methods, we selected FreeSurfer (http://surfer.nmr.mgh.harvard.edu/) (Fischl et al., 1999) and Spherical Demons (Yeo et al., 2010). FreeSurfer is the most widely used of the fully automated, surface-based, brain image analysis software packages that perform registration without requiring landmarks. FreeSurfer was recently shown to outperform affine registration and SPM Normalize (Ashburner and Friston, 1999) when applied to a pediatric population (Ghosh, personal communication). Spherical Demons is a recently introduced extension of the Demons algorithm (Thirion, 1998) to the sphere, which is reported to have comparable performance to FreeSurfer but runs at least an order of magnitude faster (Yeo et al., 2010). Other popular freely available surface registration methods include Caret (http://brainmap.wustl.edu) (Van Essen et al., 2001) and BrainVisa (http://brainvisa.info) (Cointepas et al., 2001). They were not included in this study because Caret still requires manually assigned landmarks (personal communication with Van Essen and Dierker) and we are not aware of any means to apply a nonlinear transform to an arbitrary set of labels within BrainVisa.

We conducted more than 16,000 registrations between 40 brain images, either directly to one another or via templates, with the brains represented as either volumes or as surfaces. The registration transforms were then applied to manually labeled versions of these images (on volumes and on surfaces) to evaluate registration accuracy. We used permutation tests to compare registration performance to first select the top-ranking volume and surface registration methods, and then we compared these selections with one another. The initial set from which we made our selection was the following: SyN and ART on brain images with and without skulls, SyN, FreeSurfer, and Spherical Demons via custom templates, and FreeSurfer via its default atlas.