196
M. Wu et al.
may suffer from morphological diversity of the ONH and lose important 3-D global
information.
In this chapter, we introduce an automated optic disc segmentation and C/D ratio
quantification method based on neural canal opening (NCO) detection in the optic
nerve head [24]. Unlike the methods that directly segment the optic disc in projection
image or using classifier to determine which A-scan belongs to the cup or rim, our
approach attempts to extract the NCO from SD-OCT scans for the delineation of the
disc margin. The proposed approach utilizes a two-stage strategy. The first step is to
locate the coarse disc margin by the segmentation of the retinal pigment epithelium
(RPE) layer and the smooth constraint of consecutive B-scans. In the second step,
we develop a support vector machine (SVM)-based patch search method to find the
most likely patch centered at the NCO and refine the segmentation result. Using
the NCO and reference plane, the cup border can be evaluated. Finally, the C/D
ratio is calculated by the cup diameter dividing the disc diameter. To the best of
our knowledge, our approach is the first to automatically segment the optic disc by
a NCO detection-based patching search method. The two-stage strategy combines
the global and local information for optic disc segmentation. When determining the
coarse disc margin location, we utilize the structural characteristics of ONH for initial
NCO detection, while in the patch searching procedure, the visual content similarity
near the NCO is applied for the final segmentation.
8.2.1 Overview of the Method
Figure 8.2 shows the flowchart of the proposed algorithm, which comprises two
main stages: the coarse disc margin location and the SVM-based patch search. In the
first stage, each B-scan in the volume is denoised and rescaled during preprocessing.
Then, a 3-D graph search algorithm is applied to automatically segment the RPE
layer [21]. Based on the segmentation result, we determined the initial NCO position
by the maximum curvature of the detected RPE boundary and the smooth spatial
constraint of the consecutive B-scans. In the second stage, we select image patches
from SD-OCT volumes, and utilize a probabilistic SVM classifier for training after
feature extraction of the patches. Then, the searching procedure is generated at the
region restricted by the initial NCO location. The patch of maximum probability,
centered at NCO, is regarded as the final NCO position. After the two steps, the cup
border can be calculated by the location of the NCO and the ILM boundary, and the
C/D ratio can be quantified.
M. Wu et al.
may suffer from morphological diversity of the ONH and lose important 3-D global
information.
In this chapter, we introduce an automated optic disc segmentation and C/D ratio
quantification method based on neural canal opening (NCO) detection in the optic
nerve head [24]. Unlike the methods that directly segment the optic disc in projection
image or using classifier to determine which A-scan belongs to the cup or rim, our
approach attempts to extract the NCO from SD-OCT scans for the delineation of the
disc margin. The proposed approach utilizes a two-stage strategy. The first step is to
locate the coarse disc margin by the segmentation of the retinal pigment epithelium
(RPE) layer and the smooth constraint of consecutive B-scans. In the second step,
we develop a support vector machine (SVM)-based patch search method to find the
most likely patch centered at the NCO and refine the segmentation result. Using
the NCO and reference plane, the cup border can be evaluated. Finally, the C/D
ratio is calculated by the cup diameter dividing the disc diameter. To the best of
our knowledge, our approach is the first to automatically segment the optic disc by
a NCO detection-based patching search method. The two-stage strategy combines
the global and local information for optic disc segmentation. When determining the
coarse disc margin location, we utilize the structural characteristics of ONH for initial
NCO detection, while in the patch searching procedure, the visual content similarity
near the NCO is applied for the final segmentation.
8.2.1 Overview of the Method
Figure 8.2 shows the flowchart of the proposed algorithm, which comprises two
main stages: the coarse disc margin location and the SVM-based patch search. In the
first stage, each B-scan in the volume is denoised and rescaled during preprocessing.
Then, a 3-D graph search algorithm is applied to automatically segment the RPE
layer [21]. Based on the segmentation result, we determined the initial NCO position
by the maximum curvature of the detected RPE boundary and the smooth spatial
constraint of the consecutive B-scans. In the second stage, we select image patches
from SD-OCT volumes, and utilize a probabilistic SVM classifier for training after
feature extraction of the patches. Then, the searching procedure is generated at the
region restricted by the initial NCO location. The patch of maximum probability,
centered at NCO, is regarded as the final NCO position. After the two steps, the cup
border can be calculated by the location of the NCO and the ILM boundary, and the
C/D ratio can be quantified.