| Lion's Eye
patients with
atient with a
1e disturbed
le cornea is
vith temporal
pe provides
rnea was too
eratocon has
vhereas the
( the corneal
measurement
ent,
n for a short
that suit the
a readily
must suit the
nditions will
and certainly
topography,
ative of their
sues. In the
y the success
the case of
ly necessary.
res such as
clinicians are
portant, but
ition such as
inicians will
nost critical
rmal clinical
t, namely the
€ appropriate
1ises patient
estimates of
affect vision
:quent vision
erefore meet
d for on-the-
nt,
jgnition and
] data,
vi. high quality presentation of appropriate clinical
information,
vii. near real-time presentation of data if intraoperative
applications are envisaged.
The investigations reported earlier indicated that image scale
is a critical determinant of object space accuracy. For a
typical CCD sensor, a 12mm cornea will be imaged at a
reduced scale of approximately 1:2, which could be expected
to significantly reduce accuracy. This loss may be partly
compensated for by the higher object space accuracy that can
flow from improved target centroiding, although the
relatively poor quality of corneal reference marks is likely to
limit such gains, and from the better metric performance of a
digital camera, particularly the lack of film deformation. It
would be preferable to keep the image scale close to 1:1,
which means using a camera with a CCD array that is at least
15mm x 15mm. The choice of image transfer technique is
between PLL line synchronised or pixel synchronous
framegrabbing from a CCD video camera, or direct capture
from a digital transfer camera. The latter two of these
techniques provide the best available solution for most
photogrammetric applications. An instrument that does not
allow the clinician to view the images conveniently in real-
time prior to capture would be too cumbersome for routine
clinical use. The preferred solution for this application is
therefore pixel synchronous framegrabbing of a live video
image.
In the case of the Keratocon, and for a 15mm x 15mm array
with 1024 x 1024 pixel resolution, an 80pm circular target at
scale 1:1 would be imaged onto the array as a region with a
diameter of approximately 6 pixels. A target centroiding
accuracy in the order of only 0.2 pixel would represent better
than +4jm in the image which is an improvement on the
analogue system. Higher resolution arrays would be
expected to improve this accuracy.
Figure 6. Portion of a 1:1 scale image of a targeted
cornea sampled at 3072x2048 pixels.
Target recognition is expected to be the most complicated
aspect of a digital implementation. Some experimental work
has been conducted in an attempt to ascertain the feasibility
of automatically detecting the corneal marks. A portion of a
targeted cornea imaged at 1:1 scale on 35mm film and
scanned onto a 3072x2048 array is shown in Figure 6 above.
The image has been contrast enhanced using a linear
histogram remapping but no other image processing. There
is reasonable segmentation of the dark marks from the
447
surrounding cornea and within most marks the bright spot
representing the target is well enough defined to allow
manual measurement. A control point appears as a bright
spot and is clearly segmented.
There are a number of factors that complicate automatic
recognition. The corneal reference marks are on a contact
lens that does not fit in a repeatable manner to the patient's
eye and so there is no reliable a priori knowledge of the
position of the reference marks in the image. The colour of
the iris which forms the background to the corneal images
varies greatly between patients. Further, it is difficult to
control illumination between images.
If pixel synchronous framegrabbing is used then on-the-job
calibrations would not be required. Photocontrol would only
be required for initial and regular camera calibrations. The
beamsplitter in the analogue instrument could be replaced
with a removable mirror that was fitted only for calibrations.
The limited depth of field ensures that the cameras would be
calibrated for an object space within Imm to 2mm of the
cornea's position. A magnified image of a photocontrol
target reflected from the beamsplitter but without a cornea in
the field of view is shown in Figure 7. This is again from a
3072x2048 pixel scanned image of the illuminated control
sphere. Its diameter is 16 pixels. There has been no image
processing used to segment this target from the background.
Recognising and centroiding a target such as this would pose
no problems. The targets appear in consistent positions in
the images and could be automatically identified.
Figure 7. A digital image of a control target imaged off
the beamsplitter.
5. DATA PRESENTATION AND ANALYSIS
In medical applications of photogrammetry it is particularly
important to address the issues of data presentation and
analysis. Unlike most industrial applications, the provision
of three dimensional coordinates alone is usually not
sufficient. In order for a photogrammetric measurement
system to be clinically acceptable, it must provide
information that is immediately relevant to the user and in a
form that is readily understandable. Different users (the
ophthalmologist, corneal surgeon, contact lens fitter, visual
scientist) will require different information and different
presentation (eg Missotten 1994).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
