9 Choroidal OCT Analytics
217
A. RPE inner boundary detection: Noting the higher brightness of the RPE compared to that of the adjacent layers, an initial edge map is obtained using the
gradient-based Canny edge operator [30]. However, also notice in Fig. 9.4a that
sharp change in brightness occurs not only at the RPE inner boundary, but at
the retinal inner boundary as well. Accordingly, the edge operator principally
detects both the above boundaries alongside some secondary edges (Fig. 9.4c).
Of these, the outer one of the principal edges is taken as the RPE inner boundary
and removed its discontinuities using the dilation operator (Fig. 9.4d) [30].
B. Choroid inner boundary (CIB) detection: Now the various retinal layers inside
RPE inner boundary are peeled off (Fig. 9.4e). The RPE outer boundary, which
also defines the CIB, occurs at a more or less uniform distance from the RPE inner
boundary, and is then detected based on gradient-based bright-to-dim transition.
subsequently, the RPE layer is also peeled off (Fig. 9.4f). At this point, detection
of the COB is remained. To this end, a region of interest (ROI) of sufficient
thickness outside the CIB is selected such that the ROI would contain the COB
(Fig. 9.4g).
9.2.2.2.3 Choroid Outer Boundary (COB) Detection: The steps employed for COB
detection are outlined in the flowchart of Fig. 9.3.
A. Initial COB estimate based on SSIM: Observe in Fig. 9.4a that the sclera (uniform) and the choroid (granular) have dissimilar structure. Accordingly, such
dissimilarity is exploited by taking a small window from sclera as a template,
and calculate the structural similarity (SSIM) index between the template and
the neighborhood (of the same size as the template) of every pixel throughout the
ROI. Here SSIM between two windows A and B of equal dimensions is given
by [31]
SSIM(A, B) =
(2μ A μ B + c 1 )(2σ AB + c 2 )
(2μ
2
A μ
2
B + c 1 )(σ
2
A + σ
2
B + c 2 )
,
(9.1)
where μ A and μ B denote the respective means, and σ
2
A and σ
2
B the respective
variances of windows A and B, whereas σ AB denotes their covariance. Further,
c 1 and c 2 are small constants, chosen to stabilize the expression. As the template
is chosen from the sclera, we expect scleral pixels to have higher and choroidal
pixels to have lower SSIM indices. Indeed the lowest SSIM values are observed
in the transition region between sclera and the choroid, i.e., near the COB.
More accurately, pixels with SSIM index below a suitable threshold are mostly
concentrated on the choroid side of the COB, and generally isolated on the scleral
side (Fig. 9.4h). Such scleral pixels are removed using the connected components
algorithm (Fig. 9.4i) [30]. Finally, the lower boundary of the remaining subthreshold pixels is taken as an initial estimate of the COB (Fig. 9.4j), which
however is generally discontinuous.
217
A. RPE inner boundary detection: Noting the higher brightness of the RPE compared to that of the adjacent layers, an initial edge map is obtained using the
gradient-based Canny edge operator [30]. However, also notice in Fig. 9.4a that
sharp change in brightness occurs not only at the RPE inner boundary, but at
the retinal inner boundary as well. Accordingly, the edge operator principally
detects both the above boundaries alongside some secondary edges (Fig. 9.4c).
Of these, the outer one of the principal edges is taken as the RPE inner boundary
and removed its discontinuities using the dilation operator (Fig. 9.4d) [30].
B. Choroid inner boundary (CIB) detection: Now the various retinal layers inside
RPE inner boundary are peeled off (Fig. 9.4e). The RPE outer boundary, which
also defines the CIB, occurs at a more or less uniform distance from the RPE inner
boundary, and is then detected based on gradient-based bright-to-dim transition.
subsequently, the RPE layer is also peeled off (Fig. 9.4f). At this point, detection
of the COB is remained. To this end, a region of interest (ROI) of sufficient
thickness outside the CIB is selected such that the ROI would contain the COB
(Fig. 9.4g).
9.2.2.2.3 Choroid Outer Boundary (COB) Detection: The steps employed for COB
detection are outlined in the flowchart of Fig. 9.3.
A. Initial COB estimate based on SSIM: Observe in Fig. 9.4a that the sclera (uniform) and the choroid (granular) have dissimilar structure. Accordingly, such
dissimilarity is exploited by taking a small window from sclera as a template,
and calculate the structural similarity (SSIM) index between the template and
the neighborhood (of the same size as the template) of every pixel throughout the
ROI. Here SSIM between two windows A and B of equal dimensions is given
by [31]
SSIM(A, B) =
(2μ A μ B + c 1 )(2σ AB + c 2 )
(2μ
2
A μ
2
B + c 1 )(σ
2
A + σ
2
B + c 2 )
,
(9.1)
where μ A and μ B denote the respective means, and σ
2
A and σ
2
B the respective
variances of windows A and B, whereas σ AB denotes their covariance. Further,
c 1 and c 2 are small constants, chosen to stabilize the expression. As the template
is chosen from the sclera, we expect scleral pixels to have higher and choroidal
pixels to have lower SSIM indices. Indeed the lowest SSIM values are observed
in the transition region between sclera and the choroid, i.e., near the COB.
More accurately, pixels with SSIM index below a suitable threshold are mostly
concentrated on the choroid side of the COB, and generally isolated on the scleral
side (Fig. 9.4h). Such scleral pixels are removed using the connected components
algorithm (Fig. 9.4i) [30]. Finally, the lower boundary of the remaining subthreshold pixels is taken as an initial estimate of the COB (Fig. 9.4j), which
however is generally discontinuous.