Biomedical Engineering Reference
In-Depth Information
images must have minimal interpolation errors. A closely related topic is image
registration as applied to fMRI (see Chapter 8), in which a sequence of similar
images is generally acquired at intervals of a few seconds, and changes in sub-
ject position with time are corrected by retrospective image alignment. In clini-
cal applications, a wider range of sequence types may be required and images
may be acquired weeks or months apart, rather than seconds or minutes.
The similarity of the images and the requirement of subvoxel precision
favors registration algorithms based on voxel similarity measures rather
than landmark or fiducial systems. Virtually all currently available registra-
tion algorithms using voxel similarity measures have been successfully
applied to single subject serial studies (see Chapter 3 for technical details).
Methods by Woods
1
2
and Friston
have found wide application in fMRI.
3
4
5
6
Those by Hajnal,
Lemieux,
Freeborough and Fox,
and Studholme et al.
have also been used for more medically oriented studies.
A critical issue in serial MRI studies is the threshold for detection of
change, which is the level of change that must occur for it to be detected
against background noise and artifacts. In our experience, artifacts are usually
the limiting factor in MRI. In addition to the full range of MRI acquisition
artifacts, artifacts associated directly with the registration process are impor-
tant. To control the latter it is necessary to match the data processing involved
with registration to the properties of the original MRI data and vice versa.
Serial MRI studies typically employ either (a) three-dimensional (3D) data
sets or (b) two-dimensional (2D) multislice data sets. Since the subjects
occupy three spatial dimensions, the data and registration algorithms must
also reflect this. Data requirements for 3D coverage have been discussed in
Chapter 4. While the necessary conditions are satisfied for true 3D se-
quences, this is frequently not achieved with 2D slice methods.
7
Comparing precisely registered images acquired at different times is
greatly facilitated by generating subtraction images derived from position-
ally registered scans. On these images, signals from unchanged structures
cancel out, producing a neutral background against which real differences
can be identified more clearly. These registered subtraction images require
a different approach to interpretation than conventional images.
3,8-10
This
chapter describes a formalism for analyzing them and shows examples of the
application of image registration and subtraction, as well as how quantifica-
tion can be applied to various problems.
7.2
Methods
Three-dimensional images were acquired with true 3D radiofrequency (RF)
spoiled T1-weighted pulse sequences on a 1.0T Marconi HPQ plus scanner
(Marconi Medical Systems, Cleveland, Ohio). For whole head studies, a non-
selective RF excitation pulse was employed (TR 21 msec, TE 6 msec, flip angle
35
) and images were acquired in the sagittal plane with frequency-encoding
