Page 47 - Computational Retinal Image Analysis

3 Ophthalmic instruments 37

source such as a LED, flashlamp or tungsten lamp. The color rendition is inevitably
different. SLOs may use a single wavelength or may use multiple lasers of different
wavelengths to simulate color imaging [43]. As shown in Fig. 12b, even images re-
corded with just two lasers (green, 532 nm and red, 633 nm) give a useful color effect,
largely because the reflectivity (albedo) of the retina is much higher in the red and
green and so color images of the retina contain very little blue light.
The two images of an optic disc shown in Fig. 12c, recorded with a fundus camera
and a three-wavelength SLO are however significantly different and an understanding
of these differences can be important for analysis of the images. The RGB intensities
in a color image are due to a the multiplication of the source spectrum, chromophore
absorption spectrum and the system spectral response. For conventional color imag-
ing each band is a weighted integral of the absorption spectrum, whereas for imaging
using a narrowband source such as a laser (or LED), the spectrum is sampled at only
the narrow laser wavelength. A small change in this wavelength can make a very
significant difference: for example as can be appreciated by the difference between
the images recorded at 590 nm and 600 nm in Fig. 5. In particular, note that the arter-
ies have much greater contrast at 590 nm. Differences in the color of other retinal
features, such as drusen and neural structures can also be significant [27].
The diffusion of light away from the laser spot in the retina is also important. That
is to say, the retina exhibits an extended ‘tissue point-spread function’ [11] that tends to
smooth the image in a similar way to a convolution process, although the heterogeneity
of the retina means the effect is strongly spatially variant and differs substantially from a
convolution process. The effect also depends on the size of the detector in the SLO as de-
scribed below. A detailed quantitative understanding can be obtained using Monte-Carlo
modeling, but a heuristic understanding can be quite useful as we discuss below and in
Section 2.5. As discussed above, the width of this tissue point-spread function varies
greatly between the red and near-infrared where it can be several mm to the blue where
it is a few 10 s of microns. This variation can be readily appreciated by observing the dif-
fering amounts of diffusion of a red, green or blue laser light shone onto a human hand—
light from a green or blue laser pointer diffuses only a very short distance from the laser
spot, whereas the light from a red laser will diffuse through the whole width of a thumb.
In contrast to a fundus camera, the SLO illuminates only a single point (more
precisely a diffraction-limited spot about 10 μm wide) and records scattered light
from only a restricted volume of the retina close to the illumination spot. A small
pinhole located in front of the detector, as used in a confocal SLO, restricts detected
light to the illumination ‘point’, while a larger pinhole enables detection of light from
an extended volume of retina around the illumination. In terms of a simplified model
[32] images of the vasculature (for example) recorded with a confocal SLO are due
to light that has passed twice through a blood vessel whereas for a large pinhole, light
is transmitted only once through the vessel. A small pinhole thus yields an image
with a higher contrast, but the rejection of scattered light reduces the optical intensity
and so an optimum trade of contrast against signal-to-noise ratio may be obtained by
careful empirical and modeling-based optimization of pinhole size [30]. One clear
consequence of this double-pass phenomena is that at red and infrared wavelengths