Dual-Bootstrap ICP Registration: Demo and Applications

1
Dual-Bootstrap ICP RegistrationDemo and
Applications

• Chuck Stewart
• Department of Computer Science
• Rensselaer Polytechnic Institute
• Troy, NY
• stewart_at_cs.rpi.edu

2
Feature-Based Retinal Image Registration
In retinal images the most prominent features are
the blood vessels. Our feature-extraction
algorithms pull out branching and cross-over
points (landmarks) as well as centerline points
and widths along the vessels. In the image pair
below (hit the space bar or click with the
mouse), the quality of the feature extraction is
rather poor. The images are poorly focused in
different places and the blood vessels are not
prominent. Only one landmark (space or click to
see it) is found in common between the two
images. The extracted vascular centerlines
(space or click to see them) are fragmented and
have missing / spurious points.
3
Dual-Bootstrap ICP Vascular Registration
This animation illustrates our new registration
algorithm. The algorithm starts from very small
image regions (in red) where a common landmark is
automatically guessed to occur. It attempts to
build (bootstrap) an accurate, image-wide
alignment based on this. This illustration shows
a correct guess. (Incorrect guesses are quickly
discovered and eliminated.) Extracted
vasculature from one retinal image is shown in
black and extracted vasculature from the other
image is shown in white. At the start the
alignment is good only in the small red region,
and even there it isnt great. The algorithm
gradually pulls the entire image pair into
alignment by (a) refining the alignment in the
red region, (b) expanding this region, and (c)
switching to more sophisticated transformation
models as needed. In the example shown, the
algorithm is being applied to two different
images of the same retina taken at about the same
time but with a rotation of the eye in between.
We have also applied it successfully (next two
slides) to aligning retinal images from different
modalities and images taken with time gaps of
months and years where there are substantial
structural changes. We feel this is a new
approach to registration in 2D and in 3D,
especially of elongated structures like
vasculature.
4
Changes Over Time
This animation shows three images of the same
patient taken over a 4 month period. The
animation shows images as aligned by the
Dual-Bootstrap ICP registraiton algorithm. In the
white areas on the center right (covering the
macular region), substantial changes may be seen.
Elsewhere on the vasculature, the alignment is
extremely accurate. The accuracy of this
alignment makes the true changes stand out.
5
Multi-Modal Registraton
This animation shows the alignment of two fundus
images of a patient with images from a
fluorescein angiography (FA) sequence. The two
fundus images were taken in the visible spectrum
using a ring of illumination. The FA sequence is
taken in the near IR after the injection of
fluorescein dye. In the retina, the dye moves
through the arteries, arterioles, capillaries and
into the venules very quickly. Hence, only the
first FA image in the sequence shows veins and
venules as dark. Later in the sequence, the
arteries and arterioles darken as the fluorescein
gradually filters out. The algorithm is able to
handle all of these illumination changes.
6
The technical report describing the
Dual-Bootstrap ICP algorithm may be accessed on
the web at http//www.cs.rpi.edu/stewart/papers
/dual_bootstrap_icp.ps.gz or http//www.cs.rpi.
edu/stewart/papers/dual_bootstrap_icp.pdf