[Full Chapter: 190K, .pdf]
1Paul M. Thompson PhD,
1,2Michael S. Mega MD PhD,
1Arthur W. Toga PhD
1Laboratory of Neuro Imaging, Brain Mapping Division, Department of Neurology, UCLA School of Medicine, Los Angeles, CA
2Alzheimer's Disease Center and Memory Disorders Clinic, UCLA School of Medicine
Recent developments in brain imaging have revolutionized medicine and neuroscience. The ability to image the living brain has greatly accelerated the collection and databasing of brain maps (Mazziotta et al., 1995; Huerta and Koslow, 1996; Fox, 1997; Toga and Thompson, 1998; Letovsky et al., 1998; Van Horn et al., 2001). These maps store information on anatomy and physiology, from whole-brain to molecular scales. Some capture functional changes that occur over milliseconds, and others anatomical changes occurring over entire lifetimes (see e.g. Toga and Mazziotta, 1996; Frackowiak et al., 1997, for recent reviews).
This rapid collection of brain images from healthy and diseased subjects has stimulated the development of mathematical algorithms that compare, pool and average brain data across whole populations. Brain structure is complex, and varies widely from one individual to another. New approaches in computer vision (Bankman, 1999; Fitzpatrick and Sonka, 2000; MacDonald et al., 2000; Xu et al., 1999; Shattuck and Leahy, 2001; Kriegeskorte and Goebel, 2001), anatomical modeling (Thompson et al., 2001; Fischl et al., 2001), differential geometry (Grenander and Miller, 1998; Haker et al., 1999; Hurdal et al., 1999), and statistical field theory (Friston et al., 1995; Worsley et al., 1999; Cao and Worsley, 1999; Taylor and Adler, 2001) are being formulated to capture this variation, encode it, and detect disease-specific patterns (Thompson et al., 1997, 2001). Statistics that describe how brain structure and function vary in a population can greatly empower the analysis of new images (Ashburner et al., 1997; Gee et al., 1998; Dinov et al., 2001; Kang et al., 2001). Population statistics can help algorithms find brain structures in new scans and detect abnormalities (Thompson et al., 1997, 2001; Csernansky et al., 1998; Ashburner and Friston, 2000; Narr et al., 2001; Fischl and Dale, 2001). They can also identify systematic patterns of anatomy and function (e.g. Giedd et al., 1999; Sowell et al., 1999, 2001; Paus et al., 1999; Good et al., 2001), and uncover surprising relationships between genotype and phenotype (Styner and Gerig, 2001; Thompson et al., 2001; Cannon et al., 2001).
Increased Automation. Brain mapping analyses often draw upon hundreds or even thousands of images (Evans et al., 1994; Good et al., 2001). Computational approaches must therefore distil information from these images in a highly automated way. This challenge has driven developments in automated image registration and warping methods (Woods et al., 1998; Toga, 1998; Ashburner et al., 1999; Guimond et al., 1999; Thompson et al., 2000; Shattuck and Leahy, 2001), as well as techniques for rapid image segmentation and labeling (Collins et al., 1995). As brain databases grow at a near-exponential pace, very large image analyses can now be run through client-server software pipelines. These intensive analyses may be performed on a remote server, aided by supercomputing resources to mine data for patterns and population trends (Toga et al., 2001; cf. Zijdenbos et al., 1996; Warfield et al., 1998; Megalooikonomou et al., 2000).
In this chapter, we review recent advances in image analysis that have enabled the creation of population-based brain atlases (Mazziotta et al., 1995, 2001; Thompson et al., 2000). These atlases combine imaging data from healthy and diseased populations, and have a range of exciting applications in neuroscience. They describe how the brain varies with age, gender, demographics. They also provide a comprehensive approach for studying a particular population subgroup, with a specific disease or neuropsychiatric disorder, for instance.
What do Population-Based Atlases Contain? Traditional single-subject atlases represent anatomy in a 3D coordinate system. Population-based atlases do so as well, but they contain anatomical models from many individuals. They store population averages, templates, and statistical maps, to summarize features of the population.
Central to the concept of all atlases is the idea of a common 3D reference space. The anatomy of the atlas, and new datasets that are aligned to it, can then be referred to using 3D spatial coordinates. Image registration techniques are typically used to align new datasets with the anatomic atlas (see Toga and Thompson, 2001 for a review). Multiple datasets can then be compared in a common coordinate space. Multi-modality atlases (Fig. 1) bring together and correlate brain maps from diverse imaging devices. Changes in anatomical morphology can then be related to differences in underlying neurochemistry and molecular content (Poxton et al., 1998; Mega et al., 1999). Atlases may also contain derived computational maps of various types. For instance they may contain normative statistics on anatomic variability, or on rates of brain change in development or disease. Dynamic maps of growth patterns in development are of particular interest (Lange et al., 1997; Giedd et al., 1999; Sowell et al., 1999; Thompson et al., 2001).
Relating all these structural and dynamic maps to genetic, therapeutic, and cognitive factors presents specific mathematical challenges. Among the armory of mathematical tools in building an atlas are cortical flattening approaches (Section 4), warping algorithms (Section 8), modeling approaches from population genetics (Section 9), and new results in random field theory. These mathematical tools give a multi-subject atlas its power to reveal generic patterns of brain structure and function not observable in an individual (Thompson et al., 2000).
Disease-specific brain atlases are a particular type of population atlas. They provide a unique perspective on the anatomy and physiology of a particular disease (Mega et al., 1999; Thompson et al., 2000a,b,c, 2001; Narr et al., 2001a,b; Cannon et al., 2001). These atlases store multi-modality imaging data from a specific clinical subpopulation, such as patients with Alzheimer’s disease (Mega et al., 1997, 1999), schizophrenia (Narr et al., 2001; Thompson et al., 2001), or a psychiatric disorder such as fetal alcohol syndrome (Sowell et al., 2001). The populations may be stratified to reflect a particular clinical subgroup, including those at familial risk for a disease (Cannon et al., 2001), those receiving different medications (Thompson et al., 2001), or those with a specific symptom profile or genotype.
Genetic Atlases. Inclusion of genetic data in these atlases makes it possible to go beyond describing a disease to investigate its causes. Novel mathematics can be used to mine imaging data, and to identify genetic sources of variation. This allows the direct mapping of genetic influences on brain structure, and lets us quantify heritability for different features of the brain (Thompson et al., 2001; see Section 9). Familial, twin and genetic linkage studies have recently begun to expand the atlas concept to tie together genetic and imaging studies of disease. Atlases that contain genetic brain maps, and a means to analyze them, can help screen relatives for inherited disease (Cannon et al., 2001). They also offer a framework to mine large imaging databases for risk genes and quantitative trait loci (Gottesman, 1997), as well as genetic and environmental triggers of disease.
Paul Thompson, Ph.D.
|RESUME| E-MAIL ME| PERSONAL HOMEPAGE| PROJECTS|