8 Segmentation of Optic Disc and Cup-to-Disc Ratio Quantification …
195
In [1, 6], Strouthidis et al. pointed out that neural canal opening (NCO) is an
objective anatomic landmark consistent with SD-OCT optic disc margin anatomy.
Furthermore, Hu et al. found out that the NCO can serve as longitudinally stable
reference plane and it is not likely to change with glaucomatous progression [7].
As shown in Fig. 8.1a, NCO is defined as the termination of the retinal pigment
epithelium (RPE) layer. A parallel line of 150 µm above the connecting line of
the two NCO points indicates the standard reference plane [8, 9]. The circular ring
between two intersections of the ONH surface and the reference plane are defined
as the cup border. Although there has been some debate about the determination
of the reference plane [10–12], we adopted this classic definition. In Fig. 8.2b, the
red and the green dots are the points of disc and cup margin in projection fundus
image, corresponding to the columns of the NCO and the cup borders, respectively.
Therefore, the disc margin can be constituted by the NCO in SD-OCT images.
8.2 Optic Disc Segmentation
In order to quantify the C/D ratio, many automatic methods have segmented the optic
disc and cup in color fundus photographs. For example, Aquino et al. proposed a
template-based method to detect optic disc boundary, using circular Hough transform
for boundary approximation [13]. Yu et al. also presented a hybrid level-set approach
for optic disc segmentation based on the deformable model, which combined region
and local gradient information [14]. In [15], a superpixel classification based algorithm was applied to determine the optic disc by contrast enhanced histogram and
center surround statistics. Furthermore, comparisons of the active contour models
for glaucoma screening were performed in [16].
Recently, there have been several studies for glaucoma detection in SD-OCT
images [17–23]. Antony et al. presented an automated intraretinal layer segmentation
algorithm to calculate the thickness of the retinal nerve fiber layer in normal and
glaucoma scans [17]. To detect glaucoma structural damage at an early stage, Xu
et al. generated a 2-D feature map from a SD-OCT volume by grouping super pixels
and utilized a boosting algorithm to classify glaucoma cases [18]. Another category
of approaches is based on optic disc segmentation. Work by Hu et al. [19] transformed
the SD-OCT slices to planar projection images, and used graph search algorithm to
detect the two boundaries of optic disc and cup simultaneously. In [20, 21], Lee
et al. proposed a multi-scale 3-D graph search algorithm to segment retinal surfaces
for OCT scan flattening, and then classified each voxel column (A-scan) using k-NN
classifier according to the features obtained from the projection image and the retinal
surfaces. By observing that the optic disc bounded by RPE has a different structural
appearance from the area with the disc, Fu et al. applied low-rank reconstruction to
detect the boundary of optic disc [22]. Based on this work, Miri et al. proposed a
multi-modal pixel classification method to segment the optic disc, combining stereo
fundus and SD-OCT volumes [23]. However, these previous methods required the
assistance of color fundus photographs. Additionally, A-scan-based classification
195
In [1, 6], Strouthidis et al. pointed out that neural canal opening (NCO) is an
objective anatomic landmark consistent with SD-OCT optic disc margin anatomy.
Furthermore, Hu et al. found out that the NCO can serve as longitudinally stable
reference plane and it is not likely to change with glaucomatous progression [7].
As shown in Fig. 8.1a, NCO is defined as the termination of the retinal pigment
epithelium (RPE) layer. A parallel line of 150 µm above the connecting line of
the two NCO points indicates the standard reference plane [8, 9]. The circular ring
between two intersections of the ONH surface and the reference plane are defined
as the cup border. Although there has been some debate about the determination
of the reference plane [10–12], we adopted this classic definition. In Fig. 8.2b, the
red and the green dots are the points of disc and cup margin in projection fundus
image, corresponding to the columns of the NCO and the cup borders, respectively.
Therefore, the disc margin can be constituted by the NCO in SD-OCT images.
8.2 Optic Disc Segmentation
In order to quantify the C/D ratio, many automatic methods have segmented the optic
disc and cup in color fundus photographs. For example, Aquino et al. proposed a
template-based method to detect optic disc boundary, using circular Hough transform
for boundary approximation [13]. Yu et al. also presented a hybrid level-set approach
for optic disc segmentation based on the deformable model, which combined region
and local gradient information [14]. In [15], a superpixel classification based algorithm was applied to determine the optic disc by contrast enhanced histogram and
center surround statistics. Furthermore, comparisons of the active contour models
for glaucoma screening were performed in [16].
Recently, there have been several studies for glaucoma detection in SD-OCT
images [17–23]. Antony et al. presented an automated intraretinal layer segmentation
algorithm to calculate the thickness of the retinal nerve fiber layer in normal and
glaucoma scans [17]. To detect glaucoma structural damage at an early stage, Xu
et al. generated a 2-D feature map from a SD-OCT volume by grouping super pixels
and utilized a boosting algorithm to classify glaucoma cases [18]. Another category
of approaches is based on optic disc segmentation. Work by Hu et al. [19] transformed
the SD-OCT slices to planar projection images, and used graph search algorithm to
detect the two boundaries of optic disc and cup simultaneously. In [20, 21], Lee
et al. proposed a multi-scale 3-D graph search algorithm to segment retinal surfaces
for OCT scan flattening, and then classified each voxel column (A-scan) using k-NN
classifier according to the features obtained from the projection image and the retinal
surfaces. By observing that the optic disc bounded by RPE has a different structural
appearance from the area with the disc, Fu et al. applied low-rank reconstruction to
detect the boundary of optic disc [22]. Based on this work, Miri et al. proposed a
multi-modal pixel classification method to segment the optic disc, combining stereo
fundus and SD-OCT volumes [23]. However, these previous methods required the
assistance of color fundus photographs. Additionally, A-scan-based classification