ReviewFreeSurfer
Introduction
FreeSurfer is suite of powerful tools that provide extensive and automated analysis of key features in the human brain. This includes volumetric segmentation of most macroscopically visible brain structures (Fischl et al., 2002, Fischl et al., 2004a), segmentation of hippocampal subfields (Van Leemput et al., 2009), inter-subject alignment based on cortical folding patterns (Fischl et al., 1999b), segmentation of white matter fascicles using diffusion MRI (Yendiki et al., 2008), parcellation of cortical folding patterns (Desikan et al., 2006, Destrieux et al., 2010, Fischl et al., 2004b), estimation of architectonic boundaries from in vivo data (Fischl et al., 2008, Fischl et al., 2009, Hinds et al., 2008, Yeo et al., 2009), mapping of the thickness of cortical gray matter (Fischl and Dale, 2000), and the construction of surface models of the human cerebral cortex (Dale et al., 1999, Fischl et al., 1999a).
It is this last functionality—the construction of cortical surface models—that was the motivation for the development of the software that would eventually become FreeSurfer. The surface reconstruction code traces its roots to Anders Dale's Ph.D. dissertation work with Marty Sereno in the early 1990s (Dale, 1994), in which the surface models were used to solve the EEG/MEG inverse problem (Dale, 1994, Dale et al., 2000). The EEG/MEG inverse problem is that one records electromagnetic signals outside the skull with electrodes (for EEG) or magnetometers/gradiometers (for MEG), but one wishes to recover the set of currents inside the brain that gave rise to the measured signals. This is a fundamentally ill-posed problem in that an infinite distribution of source currents can give rise to the same measurements. Thus one must apply some constraints in order to obtain a solution. In this case, Anders and Marty wanted to use the fact that pyramidal neurons in the cortex are thought to be the source of the vast majority of the EEG/MEG signal. Cortical surface models would therefore give them access to the location of the pyramidal neurons. Further, because the processes of these neurons are typically oriented perpendicular to the cortical surface, the surface models would also provide orientation constraints on the underlying dipoles. Thus, the surface models allowed a strong and neurobiologically plausible set of constraints to be applied to the inverse problem, resulting in a linear inversion that minimized the L2 norm of the solution and provided some of the first spatiotemporal movies of estimated human brain activity (Dale, 1994).
Prior to Anders’ dissertation work, others had attempted to construct surface models, as it was widely recognized how useful they would be for both visualization and analysis purposes. Most previous attempts had focused on the construction of the so-called “pial” surface, which represents the “top” of the cortical gray matter, ideally above layer I and below the pial. Unfortunately, this surface is impossible to directly visualize or reconstruct from MRI as there are many locations in the brain where adjacent banks of a sulcus are closer than 1 mm or so resolution achievable with MRI. Attempts to model this boundary directly lead to either topologically “correct” models (i.e. topologically equivalent to a sphere) that did not extend into deep sulci and thus excluded large regions of the brain, or geometrically reasonable models that included huge topological “defects”—holes and handles in the surface models. The fundamental insight that Anders and Marty had that enabled surface model construction was that the gray/white boundary, or the “bottom” of the gray matter, did not suffer from these problems given that adjacent banks are separated by at least twice the width of the gray matter, or an additional 3–7 mm. This added spacing was critical as it meant the gray/white surface could be resolved directly across the vast majority of the cortex.
The tools that came from this work provided surface models with adequate geometric accuracy that worked reasonably well with a specific and known MRI sequence, but did not constrain the topology of the surface models, nor did they provide estimates of the pial surface in addition to the gray/white boundary. They were thus of limited utility for cross-subject registration, in which having a correct topology allows one to construct an invertible map, or for cortical morphometry, which requires models of both the gray/white boundary as well as the pial surface to measure cortical thickness and volume. In addition, they required many hours of user intervention to construct each surface model, most notably for manually correcting large topological defects, which was tedious, labor intensive and had a steep learning curve.
Section snippets
Historical context
It is worth first pointing out that the history of surface-based analysis of cortical structure and function predates computer algorithms, and owes much to the pioneering work of people like David Van Essen and Eric Schwartz. David has always been one of the great proponents of “cortical cartography” and his early work with Heather Drury (Drury, 1997, Drury et al., 1996, Drury et al., 1997, Drury et al., 1998) laid the algorithmic foundation for what would become the CARET package for
Topology correction
Previous work in cortical surface reconstruction had focused on topological accuracy by deforming a surface with a known topology to lie at the specified interface in the imaging data (e.g. either gray/white or pial) (MacDonald, 1998, MacDonald et al., 1994, MacDonald et al., 2000). The issue with these models is the difficulty of generating surfaces that accurately follow the entire boundary of interest, as illustrated in Fig. 1. The left-hand image exhibits typical topological defects: a
Surface deformation and thickness estimation
The accuracy requirements on surface placement depend on what they will be used for. For example, for EEG/MEG source estimation, a surface misplacement of one or two mm will have little effect on the computed solution. Similarly, for analyzing and visualizing functional data that is typically acquired with voxels that are greater than 3 mm on each side (although this is changing!) surface accuracy is not critical. However, if one's goal is to measure morphometric changes associated with disease
Whole-brain segmentation
The surface models provided an excellent basis for the analysis of the structural and functional properties of the cerebral cortex, but not for subcortical and ventricular structures. At the time, segmentation tools were limited to a small number of tissue classes (e.g. gray matter, white matter and CSF) or subcortical structures (e.g. Louis Collins’ excellent work (Collins and Evans, 1997)). However, no tools existed that would provide a labeling of each voxel in the brain into semantically
Contributors
Many people have contributed to FreeSurfer over the years, and I would like to acknowledge them. First and foremost are Anders Dale and Marty Sereno who developed some of the initial tools, and also had many of the key technological and neuroscientific insights that made FreeSurfer possible. Doug Greve, who understands the intricacies of fMRI analysis as well as anyone in the world, is the primary author of almost all the functional analysis tools distributed with FreeSurfer (FS-FAST). Other
Conclusion
The development of a set of key technologies enabled the development of FreeSurfer, an array of image analysis tools designed to be automated, robust, accurate and relatively easy to use. This included automated geometric accurate topology correction (Fischl et al., 2001), surface-based inter-subject alignment (Fischl et al., 1999a, Fischl et al., 1999b) and whole-brain segmentation (Fischl et al., 2002, Fischl et al., 2004a). Although in many ways these still represent the core functionality
Acknowledgments
Support for this research was provided in part by the National Center for Research Resources (P41-RR14075, and the NCRR BIRN Morphometric Project BIRN002, U24 RR021382), the National Institute for Biomedical Imaging and Bioengineering (R01EB006758), the National Institute on Aging (AG022381), the National Center for Alternative Medicine (RC1 AT005728-01), the National Institute for Neurological Disorders and Stroke (R01 NS052585-01, 1R21NS072652-01, 1R01NS070963), and was made possible by the
References (103)
- et al.
Cortical surface-based analysis I: segmentation and surface reconstruction
Neuroimage
(1999) - et al.
Dynamic statistical parametric mapping: combining fMRI and MEG for high-resolution imaging of cortical activity
Neuron
(2000) - et al.
An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest
Neuroimage
(2006) - et al.
Automated MRI measures predict progression to Alzheimer's disease
Neurobiol. Aging
(2010) - et al.
Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature
Neuroimage
(2010) - et al.
Detection of cortical thickness correlates of cognitive performance: reliability across MRI scan sessions, scanners, and field strengths
Neuroimage
(2008) - et al.
Cortical surface-based analysis II: inflation, flattening, and a surface-based coordinate system
Neuroimage
(1999) - et al.
Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain
Neuron
(2002) - et al.
Sequence-independent segmentation of magnetic resonance images
Neuroimage
(2004) - et al.
Predicting the location of entorhinal cortex from MRI
Neuroimage
(2009)