Biomedical Engineering Reference
In-Depth Information
69
folding pattern of normal anatomy.
Thompson and Toga have created an
atlas of subjects with Alzheimer's Disease based on the surface registration
of manually identified curves on the cortex.
70
Work is currently ongoing to
determine which space (2D surface in 3D, flattened 2D surface, or original
3D space) is best for mapping regions of interest.
14.3
Applications
The demand for the automatic quantitative analysis in cohort studies (i.e.,
analysis of large medical-image data sets with hundreds or thousands of
scans) has been growing over the past few years, particularly in the brain
mapping community. Typical examples are the construction of probabilistic
atlases of the adult and pediatric brain,
71,72
the analysis of pathology in the
context of clinical research and clinical trials,
73
and the statistical analysis of
35
functional imaging studies.
To address the complexities of large computer
processing requirements, we have developed a production control system
(PCS) that allows the rapid implementation and parallel execution of an
analysis
a term used here to describe a sequence of processing
stages applied to a collection of input data. A pipeline consists of a sequence
of elementary operations that together make up a complex image analysis
task. The PCS ships individual processing steps (a program with its associ-
ated data) to a free CPU on the network, monitors its progress, and reports
results. This is completed for all steps in the pipeline and for all data sets in
the cohort.
At the BIC, we have developed a number of pipelines to analyze large
ensembles of data. Our typical pipeline for 3D brain image analysis includes a
number of preprocessing steps including image intensity nonuniformity cor-
rection,
pipeline,
75
and noise reduction.
76
Once preprocessing is completed, the pipeline combines linear registration
28
and resampling into stereotaxic space, cortical surface extraction,
77,78
tissue
classification,
79
automatic sulcal extraction,
80
and atlas matching using non-
linear registration.
19
Linear registration is achieved by maximization of the cross
correlation ratio between an individual volume to be registered and average
MNI305 target volume already registered to stereotaxic space.
28,36,38,39
The non-
linear registration is estimated by computing a 3D nonlinear deformation field
in a piecewise linear fashion, fitting cubical neighborhoods in sequence.
19
Each
data cube in one volume is translated to achieve an optimal match within the
other volume. Cubes are arranged in a 3D grid to fill the volume, and each cube
moves within a range defined by a grid spacing. The algorithm is applied in a
multiscale hierarchy. At each step, the image volumes are preconvolved with a
3D Gaussian kernel where blurring and cube size are reduced after each stage.
Initial fits are obtained rapidly since, at lower scales only gross distortions are
considered, but later iterations at finer scales accomodate local differences at
the price of increasing computational burden.
74
volume-to-target intensity normalization,
