Surface models of the cerebral cortex are useful in a variety of applications, including
visualization, intersubject registration and comparison, and EEG/MEG source
localization. We have developed an automated sequence that produces cerebral cortex
models with spherical topology from T1-weighted MRI of the human head. We first skullstrip
the brain using our Brain Surface Extractor (BSE). BSE uses a combination of
anisotropic diffusion filtering, Marr-Hildreth edge detection, and a sequence of
mathematical morphology operators to segment and extract the brain from the
surrounding tissues. We then compensate for tissue intensity non-uniformities using our
Bias Field Corrector (BFC). BFC computes local estimates of gain variation in the image
using an adaptive partial volume tissue model. From these estimates, BFC then
computes a tri-cubic B-spline that estimates the gain variation throughout the brain
volume. BFC corrects the image using the spline values. We next label each voxel in
the brain according to tissue content. Our tissue classifier combines our issue model
with a Gibbs spatial prior; the use of the prior encourages tissue labelings that have
contiguous regions of similar tissue types. We then extract the cerebrum from the
tissue-labeled volume by registering a labeled atlas to the skull-stripped MRI data. Isocontour
algorithms can then be used to produce a surface model from the identified
cerebrum. However, such models will typically not exhibit spherical topology. For this
reason, we apply our topological correction algorithm (TCA), which uses a graph-based
approach to ensure that the tessellations we generate will have the topology of a sphere.
Finally, we parameterize these surfaces using a method based on p-harmonic energy
minimization. In this presentation, we will describe each stage of our method and
demonstrate related software products that we have developed.
A cortical surface extracted using our BrainSuite tools, displayed in Duff.
Audio (MP3 File, Podcast Ready)