254
F. Shi et al.
had strong contrast and thus could endure image noise. Surfaces 2–6 were slightly
affected by the fovea or the PED and had weaker contrast. Therefore medium y was
set for these surfaces, with y 3 at resolution level 2, and y 6 at resolution
level 3. Surface 12 and surfaces 8–10 on the flattened image were required to be
smooth surfaces. Therefore small y was needed, set as y 1. For the refining
step in multi-resolution surface detection, assuming the initial detection was accurate
enough, the surface position in higher resolution would be close to the initial one (the
center line). Therefore, small smoothness constraints
x y 1
were applied.
For PED footprint detection, the distance thresholds and the area threshold were
selected empirically as d 1 3, d 2 1, and A 30 in pixels. However, as tested,
the region segmentation performance was not sensitive to perturbations of these
parameters. Empirically tested, the suggested ranges of parameters were: d 1 3–7,
d 2 1–2 and A 30–70. Small values were preferred so that PED regions with minor
elevation of RPE and small sizes would not be discarded. Most false positives that
were not detected by the size criteria could still be ruled out by the intensity criterion
that follows. The intensity threshold T was set as the adaptive Otsu threshold [34]
considering the fact that the OCT images have a double-peaked histogram.
10.2.3.2 Layer Segmentation Results for PED Dataset
Examples of layer segmentation results are shown in Fig. 10.8 in both 2-D and 3-D.
Table 10.2 shows the mean and standard deviation of unsigned border positioning
errors for each surface computed on 200 B-scans from the PED dataset, compared
with the inter-observer variability and the errors resulting from employing the Iowa
Reference Algorithm [14]. The p-values are shown in Table 10.3, with bold fonts
indicating that the proposed method has statistically significantly better performance.
Compared with the difference between observers, the errors of surfaces 1 and 11 are
significantly smaller, the errors of surfaces 4 and 10 are significantly bigger, and the
errors of surfaces 2, 5, 6 and 7 are statistically indistinguishable. The overall mean
unsigned error is 7.87 ± 3.36 µm, which is statistically indistinguishable from the
mean unsigned difference between observers (7.81 ± 2.56 µm). Compared with [14],
the errors of surfaces 2 and 11 are statistically significantly smaller, the errors of the
other surfaces are statistically indistinguishable, and the overall mean unsigned error
is statistically significantly smaller.
The results in Table 10.2 only show minor improvement over the method designed
for normal retina in [14] because the PED is a localized structure. Only in a small
proportion of B-scans the layers exhibit dramatic morphological changes, and in the
remaining B-scans the retina layers appears normal, so that the segmentation method
for normal retinas performed well too. To better evaluate the layer segmentation
performance near the PED, Table 10.4 shows the mean and standard deviation of
unsigned border positioning errors calculated on the 50 B-scans labeled with PED’s,
compared with both inter-observer variability and the errors resulting from employing the Iowa Reference Algorithm [14]. The p-values are shown in Table 10.5, with
bold fonts indicating that the proposed method has statistically significantly bet-
F. Shi et al.
had strong contrast and thus could endure image noise. Surfaces 2–6 were slightly
affected by the fovea or the PED and had weaker contrast. Therefore medium y was
set for these surfaces, with y 3 at resolution level 2, and y 6 at resolution
level 3. Surface 12 and surfaces 8–10 on the flattened image were required to be
smooth surfaces. Therefore small y was needed, set as y 1. For the refining
step in multi-resolution surface detection, assuming the initial detection was accurate
enough, the surface position in higher resolution would be close to the initial one (the
center line). Therefore, small smoothness constraints
x y 1
were applied.
For PED footprint detection, the distance thresholds and the area threshold were
selected empirically as d 1 3, d 2 1, and A 30 in pixels. However, as tested,
the region segmentation performance was not sensitive to perturbations of these
parameters. Empirically tested, the suggested ranges of parameters were: d 1 3–7,
d 2 1–2 and A 30–70. Small values were preferred so that PED regions with minor
elevation of RPE and small sizes would not be discarded. Most false positives that
were not detected by the size criteria could still be ruled out by the intensity criterion
that follows. The intensity threshold T was set as the adaptive Otsu threshold [34]
considering the fact that the OCT images have a double-peaked histogram.
10.2.3.2 Layer Segmentation Results for PED Dataset
Examples of layer segmentation results are shown in Fig. 10.8 in both 2-D and 3-D.
Table 10.2 shows the mean and standard deviation of unsigned border positioning
errors for each surface computed on 200 B-scans from the PED dataset, compared
with the inter-observer variability and the errors resulting from employing the Iowa
Reference Algorithm [14]. The p-values are shown in Table 10.3, with bold fonts
indicating that the proposed method has statistically significantly better performance.
Compared with the difference between observers, the errors of surfaces 1 and 11 are
significantly smaller, the errors of surfaces 4 and 10 are significantly bigger, and the
errors of surfaces 2, 5, 6 and 7 are statistically indistinguishable. The overall mean
unsigned error is 7.87 ± 3.36 µm, which is statistically indistinguishable from the
mean unsigned difference between observers (7.81 ± 2.56 µm). Compared with [14],
the errors of surfaces 2 and 11 are statistically significantly smaller, the errors of the
other surfaces are statistically indistinguishable, and the overall mean unsigned error
is statistically significantly smaller.
The results in Table 10.2 only show minor improvement over the method designed
for normal retina in [14] because the PED is a localized structure. Only in a small
proportion of B-scans the layers exhibit dramatic morphological changes, and in the
remaining B-scans the retina layers appears normal, so that the segmentation method
for normal retinas performed well too. To better evaluate the layer segmentation
performance near the PED, Table 10.4 shows the mean and standard deviation of
unsigned border positioning errors calculated on the 50 B-scans labeled with PED’s,
compared with both inter-observer variability and the errors resulting from employing the Iowa Reference Algorithm [14]. The p-values are shown in Table 10.5, with
bold fonts indicating that the proposed method has statistically significantly bet-