A four-dimensional probabilistic atlas of the human brain

Abstract

The authors describe the development of a four-dimensional atlas and reference system that includes both macroscopic and microscopic information on structure and function of the human brain in persons between the ages of 18 and 90 years. Given the presumed large but previously unquantified degree of structural and functional variance among normal persons in the human population, the basis for this atlas and reference system is probabilistic. Through the efforts of the International Consortium for Brain Mapping (ICBM), 7,000 subjects will be included in the initial phase of database and atlas development. For each subject, detailed demographic, clinical, behavioral, and imaging information is being collected. In addition, 5,800 subjects will contribute DNA for the purpose of determining genotype- phenotype-behavioral correlations. The process of developing the strategies, algorithms, data collection methods, validation approaches, database structures, and distribution of results is described in this report. Examples of applications of the approach are described for the normal brain in both adults and children as well as in patients with schizophrenia. This project should provide new insights into the relationship between microscopic and macroscopic structure and function in the human brain and should have important implications in basic neuroscience, clinical diagnostics, and cerebral disorders.

Figure 1

The effect of the rapid growth in neuroscience information on the field. Left, As the amount of information generated by neuroscientists, and human neuroimaging data in particular, increases at exponential rates, so does specialization within the field. Right, Concomitant with specialization is a divergent and isolating trend that generates sub-subspecialty journals, meetings, and opportunities for information exchange. The goal of informatics projects, with the probabilistic human brain atlas as an example, is to integrate information across isolated niches and to provide a convergence of these data sets in a form that allows easy, practical, automated, and quantifiable cross-correlations between information sets using the brain itself as the spatial reference point for location and the age of an individual subject as the time reference. Crosscorrelations between species would require appropriate anatomic homologs to be developed among different species.

Figure 1

The effect of the rapid growth in neuroscience information on the field. Left, As the amount of information generated by neuroscientists, and human neuroimaging data in particular, increases at exponential rates, so does specialization within the field. Right, Concomitant with specialization is a divergent and isolating trend that generates sub-subspecialty journals, meetings, and opportunities for information exchange. The goal of informatics projects, with the probabilistic human brain atlas as an example, is to integrate information across isolated niches and to provide a convergence of these data sets in a form that allows easy, practical, automated, and quantifiable cross-correlations between information sets using the brain itself as the spatial reference point for location and the age of an individual subject as the time reference. Crosscorrelations between species would require appropriate anatomic homologs to be developed among different species.

Figure 2

The magnitude of neuroinformatics data for the human brain. Although this illustration is based on a number of assumptions, the orders of magnitude are realistic and enormous. They depict what would be involved in developing an organized data structure that combines location in the human brain with gene expression maps. Left, This example assumes that approximately 50,000 genes may be expressed in any three-dimensional region (voxel) of the brain at any given time during development. The volume of the typical human male brain is 1,500 cc. Depending on the spatial resolution used to determine gene expression (ranging from 1 cc to 10³μ³), the number of data points ranges from 75 million to 75 thousand trillion. Keep in mind that this is what is required to do just one brain at a given point in time. Right, If the same assumptions and range of resolutions are used, the range of data magnitudes for a series of brains collected across a population,with a representation of each age from birth to age 100 years, results in data set magnitudes that range from 10⁹ to 10²³. These truly astronomic orders of magnitude will require innovative, practical neuroinformatic data structures that allow the referencing of such information as a function of both location and time.

Figure 2

The magnitude of neuroinformatics data for the human brain. Although this illustration is based on a number of assumptions, the orders of magnitude are realistic and enormous. They depict what would be involved in developing an organized data structure that combines location in the human brain with gene expression maps. Left, This example assumes that approximately 50,000 genes may be expressed in any three-dimensional region (voxel) of the brain at any given time during development. The volume of the typical human male brain is 1,500 cc. Depending on the spatial resolution used to determine gene expression (ranging from 1 cc to 10³μ³), the number of data points ranges from 75 million to 75 thousand trillion. Keep in mind that this is what is required to do just one brain at a given point in time. Right, If the same assumptions and range of resolutions are used, the range of data magnitudes for a series of brains collected across a population,with a representation of each age from birth to age 100 years, results in data set magnitudes that range from 10⁹ to 10²³. These truly astronomic orders of magnitude will require innovative, practical neuroinformatic data structures that allow the referencing of such information as a function of both location and time.

Figure 3

Human visual area V5. A, Bilateral PET-CBF (positron emission tomography–cerebral blood flow) images showing activation of V5 in four separate subjects. The V5 images are superimposed on the subjects' structural MRI studies. Notice the consistent relationship between the activated site and the ascending limb of the inferior temporal sulcus. (The activated sites are shown here in black, for visibility, but are apparent as shades of red on the original published images.) B, This area also coincides with the cortical region (arrow) identified by Flechsig in 1920 as being myelinated at birth. (A and B are reprinted, with permission, from Watson JD, Myers R, Frackowiak RS, et al. Area V5 of the human brain from a combined study using positron emission tomography and magnetic resonance imaging. Cereb Cortex. 1993;3:79–94. Copyright © 1993 Oxford University Press.) C, Brain of patient studied by Zihl et al.,^, with damage to the V5 area resulting in a selective disturbance of visual motion perception.

Figure 4

Strategy for use of morphometric data in a phenotype–genotype experiment. If this distribution were for hippocampal volume, for example, candidate genes for hippocampal size (e.g., Apo E ɛ2, ɛ3, and ɛ4 alleles) could be tested against human imaging data. Human heritability can be evaluated with human twin data collected in this project. Then, extremes (e.g., the top and bottom 5 percent) of hippocampal volumes could be assessed against the candidate genes for the effect they exert on human hippocampal volume. For example, hippocampal atrophy is invariably found in Alzheimer's disease. Patients with Alzheimer's disease also have a higher probability of the genotype Apo E ɛ4 (vs. ɛ2 or ɛ3). Whether 20- to 40-year-olds have a correlation between hippocampal volume and the Apo E genotype could be determined by use of atlas data in this fashion.

Figure 5

Coronal image showing muscarinic receptors labeled with a tritiated ligand from one hemisphere of a cryosectioned brain and showing the anatomic detail that such chemoarchitectural maps can provide. When serial sections are obtained and stained for a wide range of receptors, anatomic features, and gene expression maps, a tremendous wealth of information is available for comparison with sites of functional activation obtained using in vivo techniques and macroscopic brain structure (gyri, sulci, deep nuclei, white matter tracts). Having a probabilistic strategy for relating these different types of anatomic features will provide new insights into the relationship of structure and function on both microscopic and macroscopic levels for the human brain and, by analogy, the brains of other species. The analysis of the regional and laminar distribution patterns of transmitter receptors is a powerful tool for revealing the architectonic organization of the human cerebral cortex. The authors succeeded in preparing extra-large serial cryostat sections through an unfixed and deep-frozen human hemisphere. Neighboring sections were incubated with tritiated ligands for the demonstration of 15 different receptors of all classical transmitter systems; this image shows, as an example, the distribution of [³H]oxotremorine-M binding to chorgic muscarinic M₂ receptors. Receptor autoradiographs permit the distinction of numerous borders of cortical areas and subcortical nuclei by localized changes in receptor density and regional/laminar patterns. For example, the M₂ receptor subtype clearly labels the primary sensory cortices (at the level of the section shown in the figure, e.g., the primary somatosensory area BA3b and the primary auditory area BA41) by very high receptor densities sharply restricted to both areas. The different receptors allow the multimodal molecular characterization of each area or nucleus by the so-called receptor fingerprint typing. A receptor fingerprint of a brain region consists of a polar plot based on the mean density of each receptor in the same architectonic unit (area, nucleus, layer, module, striosome, etc.). The following areas and nuclei can be delineated in the present example—cingulate cortex, motor cortex, primary somatosensory cortex, inferior parietal cortex, insular cortex, primary auditory cortex (BA41), non-primary auditory cortex, inferior temporal association cortex, entorhinal cortex, mediodorsal thalamic nucleus, and putamen. (K. Zilles , A. Toga, N. Palomero-Gallagher, and J. Mazziotta, unpublished observation.)

Figure 6

A, Autosegmentation of structures. This image illustrates the first stage of autosegmentation once the brain has been spatially normalized. Lobes, gyri, and some subcortical nuclei are labeled. This iterative process continues with increasing refinement. B, Three-dimensional model with autosegmented ventricular system. This model shows an autosegmented ventricular system converted to a surface model, enabling morphometric statistics to be calculated. This segmentation was the result of a combination of tissue classification approaches and template matching following spatial normalization.

Plate 1

(Opposite, top) Location and extent of Broca's region (Brodmann's areas 44 and 45), as defined in serial coronal sections of an individual brain after three-dimensional reconstruction; lateral views of the left hemisphere are shown. Probability maps of Broca's region, based on microscopic analysis of ten human brains, can be referenced, also in a probabilistic fashion, to functional activation sites associated with the functions of Broca's area, using the multimodality probabilistic atlas strategy. The overlap of individual postmortem brains is color-coded for each voxel of the reference brain (color bar); for example, seven of ten brains overlapped in the yellow-marked voxels.

Plate 2

(Opposite, middle) Frontal matter: automated vs. manual image segmentation. A group of expert neuroanatomists labeled a template brain using a rigorously developed set of rules and manual segmentation methods and using in vivo MRI studies and postmortem cryomacrotome data sets. Linear and nonar warping algorithms were then used to match brains obtained in this program to the template in such a way that each voxel acquires the template brain's label for that brain region. The brain is then reverse-transformed back to its native shape and entered into the probabilistic database. To test the relative accuracy of this process, in vivo MRI studies obtained in this project from 10 individual brains were distributed to three of the participating institutions. Neuroanatomists at each site used the rules that were employed to prepare the template brain to manually segment the 10 newly acquired studies. The composite accuracy of these manually segmented data sets, performed independently in triplicate at the three institutions, were compared with the performance of the automated process. Significant differences in the labeling of voxels is indicated in the panel on the right. As you can see, the automated process performs comparably with the very labor-intensive strategy required to manually segment the brains. Also notice that the region of the brain chosen for this validation was the frontal cortex, a portion of the brain known to have significant variability among human beings. This strategy thus allows for the automated segmentation of brains, thereby making possible detailed data analysis in populations large enough to develop probabilistic information about human brain structure on a macroscopic scale.

Plate 3

(Opposite, bottom) Surface models. Three-dimensional models can be created to represent major structural and functional interfaces in the brain. a, A model of the lateral ventricles, in which each element is a three-dimensional parametric surface mesh. b and c, Average ventricular models from a group of patients with Alzheimer's disease (N = 10) and matched elderly controls (N = 10). Notice the larger ventricles in the patients and a prominent ventricular asymmetry (left larger than right). This feature emerges only after averaging models for a group of subjects. Population average maps of cortical anatomy (d) reveal a clear asymmetry in perisylvian cortex. e, The cortex from an individual brain (brown mesh) overlaid on an average cortical model for a group. f, Differences in cortical patterns are encoded by computing a three-dimensional elastic deformation (pink indicates large deformation) that reconfigures the average cortex into the shape of the individual, matching elements of the gyral pattern exactly. These deformation fields store detailed information on individual deviation, and can be averaged across subjects to create three-dimensional variability maps, revealing fundamental patterns of anatomic variability in the brain (g). Tensor maps (h, color ellipsoids) reveal the directions in which anatomic variation is greatest. The ellipsoids are more elongated in the directions in which structures tend to vary the most; pink denotes largest variation; blue, least. These statistical data can be used to detect patterns of abnormal anatomy in new subjects. (Panels e through h are reprinted, with permission, from Thompson PM, Woods RP, Mega MS, Toga AW. Mathematical/computational challenges in creating deformable and probabilistic atlases of the human brain. Hum Brain Mapp. 2000;9(2):81–92. Copyright © John Wiley & Sons, 2000.)

Plate 4

(Opposite, top left) Probabilistic adult human brain: surface-rendering of the probabilistic atlas (N = 100) thresholded to 40 percent to generate a regional probabilistic isocontour for the individual brain regions on the dorsolateral surface. Probabilistic isocontours and confidence limits can be arbitrarily established for any brain region at any probabilistic level, thereby giving a sharply demarcated boundary surface for the region of interest.

Plate 5

(Opposite, top right) Age-related changes in white-matter density in the internal capsule (left) and the left arcuate fasciculus (right). The thresholded maps of t-statistic values (t<4.0) are superimposed on axial (capsule) and sagittal (arcuate) sections through the magnetic resonance image of a single subject. The images depict the exact brain locations that showed statistically significant correlations between white matter density and the age of the subject (n=111; age range, 4–17 years). The internal capsule contains fibers that carry nerve impulses from the motor cortex to the spinal cord and, eventually, to the hand muscles. The arcuate fasciculus contains fibers connecting the posterior (Wernicke's) and anterior (Broca's) speech areas of the left hemisphere.

Plate 6

(Opposite, middle) Mapping growth patterns in children. Growth rates are mapped for the corpus callosum (a), the major fiber tract that communicates information between the two brain hemispheres. The maps are based on scans obtained from the same child at ages 3 and 6 years, from another child at ages 6 and 7 years, and so on. Extremely high growth rates (up to 80 percent gain of tissue locally) can be seen in specific brain regions. Fastest growth (red) is found consistently, across ages 6 to 13 years, in the callosal isthmus, which carries fibers to areas of the cerebral cortex that support language function and areas of the temporoparietal cortex that support mathematical thinking. Growth rates in the fibers projecting to language cortex are dramatically reduced after puberty (11 to 15 years). Notice how different the peak growth rates are in a child between the ages of 3 and 6 years, where 80 percent growth occurs in frontal regions that support the planning of new actions and the organization of new behaviors. Brain tissue is also lost during development. A rapid loss of tissue in the caudate nucleus of a 7- to 11-year-old child is shown in panels b through d. This structure supports learned motor behavior. Loss of tissue may suggest localized increases in processing efficiency and elimination of redundant brain tissue as development progresses. These growth patterns are complex, as rapid growth is also occurring close to the site where tissue is lost. (Reprinted, with permission, from Thompson PM, Giedd JN, Woods RP, MacDonald D, Evans AC, Toga AW. Growth patterns in the developing brain detected by using continuum mechanical tensor maps. Nature. 2000;404(6774):190–3. Copyright © 2000, Macmillan Magazines Ltd.)

Plate 7

(Opposite, bottom) Variance of frontal cortex in normal subjects vs. patients with schizophrenia, by gender. Anatomic variability in the frontal cortex is far greater in patients with schizophrenia than in control subjects matched for age, gender, and other demographic factors. This indicates an aberrant organization of the gyral pattern in frontal cortex, perhaps occurring during late embryonic development, when the gyral pattern of frontal cortex is established. Notice that the pattern of greater anatomic variability is specific to frontal cortex and is found in both male and female patients (SZ) but not in normal controls (NC). (Reprinted, with permission, from Narr K, Thompson P, Sharma T, et al. Three-dimensional mapping of gyral shape and cortical surface asymmetries in schizophrenia: gender effects. Am J Psychiatry. 2001;158(2):244–55. Copyright © 2001 by American Psychiatric Association, Inc.)