Biomedical Engineering Reference
In-Depth Information
both video images. Minimizing this cost function to solve for the pose of the
modeled object gave an accurate (1.27 mm) and robust registration for four
images of a volunteer's face.
Video can provide tracking and registration without attaching any device to
the patient, though this requires a large area of skin to be visible. As a result,
the method is probably most applicable to initial registration for surgery
before draping, or for radiotherapy.
12.4.2
X-ray Fluoroscopy
A fluoroscopy set provides real-time x-ray images of the patient. The pers-
pective geometry is the same as that of a camera, except the patient is placed
between the pinhole (x-ray source) and the imaging plane. Also, x-rays
project through the patient, so fluoroscopy images contain information about
internal structure in particular bony detail.
12.4.2.1
Calibration
The projection geometry of an x-ray set is essentially the same as that of a
video camera and hence the methods of calibration are the same. The only
difference is that the 3D calibration object must have x-ray visible markers,
usually lead or aluminium balls, placed in known relative positions. These
positions can be defined during manufacture or measured using a CT scan.
12.4.2.2
Registration
Several methods for registration of a 3D preoperative scan to the 2D x-ray
projection have been proposed. Contour-based algorithms have been pro-
posed in which the projection of a segmented surface from the preoperative
scan is matched to the outline of the same structure in the x-ray image.
33-35
These algorithms tend to use physical measurements, such as the mean dis-
tance between the 2D x-ray projection and a projection of the 3D segmented
surface, as a similarity measure. Such algorithms are efficient to run after fea-
ture extraction has occurred. However, automatic, fast, and accurate feature ex-
traction from a complex scene, such as an interventional fluoroscopy image, is
a difficulty task
35
and the final registration result is susceptible to errors in
segmentation.
34,36
Algorithms based on image intensity require little or no feature extraction
but rely on production of digitally reconstructed radiographs (DRRs) from
the preoperative CT scan.
36,37
The similarity measures used by these algo-
rithms compare the pixel intensities in the fluoroscopy image and the DRR.
They are usually statistically based, e.g., cross correlation.
37
Recent devel-
opment of the similarity measure has resulted in cost functions such as pat-
tern intensity
38
or gradient difference,
39
that have been shown to be robust
to both low frequency intensity gradients caused by overlying soft tissue
structures and the presence of interventional instruments such as a catheter
