NeuroImage 11, Number 5, 2000, Part 2 of 2 Parts 10 E al” METHODS - ACQUISITION Construction, Testing, and Validation of a Sub-Volume Probabilistic Human Brain Atlas for the Elderly and Demented Populations Michael S. Mega*?, Christine Fennema-Notestine+, Ivo D. Dinov*, Paul M. Thompson*‘, Sarah L. Archibald.& Christopher Lindshield”, Jenaro Felix*?, Arthur W. Toga*, Terry L. Jernigani *Laboratory of New-o Imaging, Department of Neurology, UCLA School of Medicine, Los Angeles, CA ?Alzheimer Disease Research Center, Depurtment of Neurology, UCLA School of Medicine, Los Angeles, CA *Brain Image Analysis LaboratoT, UCSD School of Medicine, San Diego, CA Objective: We develop, evaluate and validate a new complex and fully stochastic anatomic sub-volume probabilistic atlas (SVPA) for the elderly and demented human brain. Background: Recent developments in human brain mapping’ have instigated the extension of average deterministic anatomic brain atlases to a new class of probabilistic sub-volume population specific atlases. Aging and dementing diseases impose the most extreme changes on brain structure and function making existing atlases inappropriate for use in these populations. Disease specific SVPA’s overcome this structural mismatch. After construction of a continuum mechanical average brain atlas’ manually outlined regions of interest on individual brains, registered to that atlas, can produce stochastic probabilistic sub-volumes encoding anatomic, registration, and functional variability. Methods: Double echo MRI scans of 10 normal elderly and 10 age-matched probable Alzheimer’s disease (AD) subjects were manually segmented into 2 hemispheres each of 30 sub-volumes of interest (SVIs) with 3 different tissue types’ and then registered to each subject’s high resolution SPGR data obtained at the same scanning session. For each of the 180 SVIs in SPGR native space we consbucted two stereotactic probabilistic representations-a linear-SVPA based on 12 parameter registration4 to the Atlas and a nonlinear-SVPA based on 6th order warping4-both reflecting the chance a subject’s SVI occurs at each voxel location in the Atlas across the 20 subjects. Validity of the linear and nonlinear SVPAs was evaluated using the native SPGR data following application of a semi-automated minimum distance classification algorithmS used to segment the scans into WM, GM, CSF and background voxels after inhomogeneity correction was accomplished. Three different validation schemes were applied to determine the best approach for volumetric analysis of brain anatomy using the SVPA-Atlas. Scheme A measures the counts of each SPGR derived WM, GM and CSF classification within the linellr-SVPA. Scheme B measures the same quantities by registering the Atlas (via a 12 parameter transform) into each individual’s SPGR native space. While, Scheme C measures the tissue counts by warping the nonlinear-SVPA Atlas (via a 6th order warp’) to each of the SPGR studies after a 7-parameter (rigid body plus a uniform scale) registration to the Atlas corrected for alignment and head size difference. The tissue segmentation counts (BGM in GM + CSF compartments) from each of the 3 schemes were compared to the native space counts for each SVI to select the best registration strategy and space for future application. Results: Correlation values relating the native space counts to the Atlas derived counts across subjects’ and SVIs’ mean values identified schemes A (r2 = 0.983) and C (r2 = 0.98) as superior over B (r’ = 0.977). Of 60 total regions 27 were significantly atrophied in AD according to native space counts. Scheme C selected 16 of these SVIs as significantly atrophied compared to 18 SVIs selected by Scheme A. The greatest magnitude of significance across SVls was in the linear, compared to nonlinear, approach indicating superior sensitivity. Conclusion: A deformable high resolution sub-volume probabilistic atlas. appropriate for structural and functional imaging analysis of Ihc elderly and demented populations, is now available. References: I. Mazziotta JC, Toga AW, Evans AC, et al. A probabilistic atlaa of the human brain: theory and rationale for it\ development. Neuroimage 1995;2:89-101. 7. Thompson PM, Woods RP, Mega MS, Toga AW. Mathematical and computational challenges in creating dcformablc and probabilistic atlases of the human brain. Human Brain Mapping 2000; (in press). 3. Jemigan TL, Ostergaard AL, Fennema-Notestine C. Mesial temporal, diencephalic, and striatal contributions to single word reading. word priming, and recognition memory. Journal of the international Neuropsychological Society 2000: (in press). 4. Woods RP, Grafton ST, Watson JDG, et al. Automated image registration: II. Intersubject validation of linear and nonlinear models. J Comput Assist Tomogr 1998;22:153-165. 5. Kollokian V. Performance analysis of automatic techniques for tissue classification of magnetic resonance images of the human brain [Masters]. Concordia University, Department of Computer Science, Montreal, Canada, 1996. Support for this work was provided by an NIA career development award (KOI(AGlOO784) to MSM; an NIA Alzheimer’s Disease Research Center grant (P50 AG16570). Alzheimer’s Disease Resources Center of California grant, the Sidell-Kagan Foundation. the Human Brain Project (NIMH/NIDA: P20MH/DA 52176, NSF (BIR9322434), and NCRR (RR05956). s597