Detecting 3D Patterns of Alzheimer's Disease Pathology with a
Probabilistic Brain Atlas based on Continuum Mechanics

Objective. We report the construction of the first probabilistic atlas of the human brain to represent a patient population.

Design/Methods. In this study, patterns of 3-dimensional variability, based on 1680 structures, were encoded and used to detect pathology in new patients. Using a hierarchical combination of two non-linear image registration algorithms [2,3], comprehensive probabilistic atlases of the brain in normal aging and Alzheimer's Disease (AD) were created from a reference archive of T1-weighted 256x256x170 resolution 3D MRI scans of 10 AD patients (age: 71.9+/-10.9 yrs.) and 10 controls matched for age (72.9+/-5.6 yrs.), gender and handedness.

Connected systems of parametric surface meshes [4,5,6] were used to model 84 structures per brain: 16 deep sulcal, callosal and hippocampal surfaces, all major cortical sulci, Sylvian fissures, 14 ventricular regions, and 36 gyral and cytoarchitectural boundaries in 3 dimensions. 3D patterns of structural variability and asymmetry were analyzed by digitally mapping all 1680 structures into a variety of stereotaxic systems, using Talairach, affine, and automated polynomial mappings [2] of increasing order. Finally, high-dimensional volumetric maps (0.1 billion degrees of freedom) were computed [3], fluidly reconfiguring the anatomy of different subjects into structural correspondence. Resulting information on variations in gyral and subcortical topography was encoded as a non-stationary Gaussian random tensor field [5] and used to detect structural anomalies in new subjects. Fluid matching of cortical patterns [6] also enabled the construction of an average surface representation of the cortex for each group, and an average intensity image template was constructed for the AD group and for controls.

Results. In the Alzheimer's patient group, regionally-selective fiber loss at the isthmus of the corpus callosum (p < 0.025) matched severe increases in left temporo-parietal and occipital horn variability, suggesting left greater than right degeneration in early AD. AD also significantly increased normal Sylvian fissure and ventricular asymmetries (p < 0.0002). Digital transformations reduced inter-subject anatomic variability from peak 3D r.m.s. values of 20 mm in Talairach stereotaxic space to 12 mm after 8th order polynomial transformation, and near-zero values after continuum-mechanical transformation.

Conclusions. Population-based brain atlasing systems offer a framework to (1) create maps of group-specific variations in anatomy, (2) detect abnormal structure in new patients, and (3) factor out confounding anatomical variance to any degree desired in functional imaging applications.

References. [1] Mazziotta JC et al. (1995) NeuroImage 2:89-101; [2] Woods et al. (1998) J. Comp. Assist. Tomography 22:155-165; [3]-[6]: Thompson PM et al.: [3]. IEEE Transactions on Medical Imaging (1996) 15(4):1-16. [4] Journal of Neuroscience (1996) 16(13):4261-4274; [5]. NeuroImage (1996) 3(1):19-34; [6]. J. Comp. Assist. Tomography 21(4):567-581.

Grant Support. (PT:) Howard Hughes Med. Inst., United States Information Agency, US-UK Fulbright Commission; (MSM:) NIA Grant K08AG100784; (AWT:) P20MH/DA52176, BIR9322434, LM/MH05639, RR05956.

Related Publications

Contact Information