Biomedical Engineering Reference
In-Depth Information
variability is large when compared to the size of a subtle signal difference that
is to be detected.
Two basic paradigms have been proposed in the literature to replace man-
ual segmentation methods and automate the analysis process, making it as
objective as possible. The first is based on segmentation and the second is
based on registration.
There have been a number of attempts at developing semiautomated
8-15
segmentation procedures that are able to identify the borders of structures of
interest, with differing levels of success for different tasks. These methods
require some level of manual intervention to align predefined atlas structures
onto the volumetric image data. A hierarchical system is normally used,
where the entire atlas is fitted with a global transformation, followed by cus-
tomization of individual atlas structures. Fully automated techniques
attempt to match atlas structures to the image data directly, thus obviating
the need for user input, with different levels of success.
16-23
Registration-based comparison techniques examine the data on a voxel-by-
voxel basis, and thus require data to be placed into correspondence spatially
to replace the aforementioned structure-by-structure comparison tech-
niques.
24-29
These techniques have the advantage of permitting exploratory
analysis of the whole brain volume without specifically identifying particu-
lar regions in each volume. The first half of this chapter will concentrate on
the technical aspects of anatomical registration. The second half will summa-
rize a number of example cohort studies that have been carried out in the BIC
at the MNI.
14.2
Technical Issues
Registration-based comparison methods are based on spatial normalization
where a spatial transformation is found to map similar structures (or homol-
ogous points) from different data sets to same spatial position. While the gen-
eral case must concern itself with the mapping of data from different imaging
modalities (intermodality registration), we deal here only with the specific
case of intramodality registration of MRI data.
Four aspects of the normalization procedure must be identified and well
defined before continuing:
•
Reference space:
Which data set and coordinate system defines the
reference frame used for comparisons?
•
Spatial mapping function:
How will brain data from a given subject
be mapped or transformed from the native (original scan) representation
to the reference frame?
•
Similarity measure:
How will a brain volume be compared to the target
data set to determine value (goodness) of a given mapping function?
