8 Segmentation of Optic Disc and Cup-to-Disc Ratio Quantification …
197
8.2.2 Coarse Disc Margin Location
The purpose of this step is to limit NCO region for the subsequent patch search
procedure. This step involves image preprocessing, RPE segmentation and initial
NCO detection.
8.2.2.1 Preprocessing
Since coherent optical beams have low temporal coherence and high spatial coherence, SD-OCT images contain speckle noise [25]. Speckle size may vary in the
axial and lateral dimensions, which is mainly determined by source bandwidth and
numerical aperture. In addition, shot noise is also present in SD-OCT images, which
can be adequately described by the additive white Gaussian noise (AWGN) process
[26]. To facilitate the segmentation of RPE and ILM layer, we pre-processed the
images by bilateral filtering denoising [27, 28], which has proven adequate for edge
preservation.
For added effectiveness in the NCO detection, we also resampled the SD-OCT
images in x-axis direction. Specifically, each B-scan in the volume was rescaled from
200 × 1024 to 600 × 400 in the axial and x-axis direction, respectively. An example
of a B-scan after denoising and rescaling is shown in Fig. 8.3b.
8.2.2.2 RPE Layer Segmentation
The 3-D graph search-based segmentation algorithm has proven to be effective for
detecting multiple intraretinal surfaces [29–31]. With some appropriate feasibility
constraints, the surface segmentation problem can be directly converted to a minimum
closed set finding problem in a geometric graph and solved in a low-order polynomial
time. We applied the multi-scale 3-D graph search approach introduced in [14] to
segment the outer boundary of the RPE for subsequent NCO detection. As shown
in Fig. 8.3c, the green and yellow curves indicate the segmented ILM and RPE
boundaries, respectively.
8.2.2.3 Initial NCO Detection
From the RPE curve fitting in Fig. 8.3c, we observed that the visible parts of the RPE
follow an approximately linear curvature, and change dramatically at the termination
of the RPE. Therefore, we selected points with maximum curvature as the NCO.
However, there are several morphological characteristics that should be noticed: (1)
The NCO points are restricted in a certain region by the imaging position and prior
anatomical knowledge. The NCO always localizes near the center of the image, as
the images are taken aimed to be centered at the optic disc, and it is also noticeable
197
8.2.2 Coarse Disc Margin Location
The purpose of this step is to limit NCO region for the subsequent patch search
procedure. This step involves image preprocessing, RPE segmentation and initial
NCO detection.
8.2.2.1 Preprocessing
Since coherent optical beams have low temporal coherence and high spatial coherence, SD-OCT images contain speckle noise [25]. Speckle size may vary in the
axial and lateral dimensions, which is mainly determined by source bandwidth and
numerical aperture. In addition, shot noise is also present in SD-OCT images, which
can be adequately described by the additive white Gaussian noise (AWGN) process
[26]. To facilitate the segmentation of RPE and ILM layer, we pre-processed the
images by bilateral filtering denoising [27, 28], which has proven adequate for edge
preservation.
For added effectiveness in the NCO detection, we also resampled the SD-OCT
images in x-axis direction. Specifically, each B-scan in the volume was rescaled from
200 × 1024 to 600 × 400 in the axial and x-axis direction, respectively. An example
of a B-scan after denoising and rescaling is shown in Fig. 8.3b.
8.2.2.2 RPE Layer Segmentation
The 3-D graph search-based segmentation algorithm has proven to be effective for
detecting multiple intraretinal surfaces [29–31]. With some appropriate feasibility
constraints, the surface segmentation problem can be directly converted to a minimum
closed set finding problem in a geometric graph and solved in a low-order polynomial
time. We applied the multi-scale 3-D graph search approach introduced in [14] to
segment the outer boundary of the RPE for subsequent NCO detection. As shown
in Fig. 8.3c, the green and yellow curves indicate the segmented ILM and RPE
boundaries, respectively.
8.2.2.3 Initial NCO Detection
From the RPE curve fitting in Fig. 8.3c, we observed that the visible parts of the RPE
follow an approximately linear curvature, and change dramatically at the termination
of the RPE. Therefore, we selected points with maximum curvature as the NCO.
However, there are several morphological characteristics that should be noticed: (1)
The NCO points are restricted in a certain region by the imaging position and prior
anatomical knowledge. The NCO always localizes near the center of the image, as
the images are taken aimed to be centered at the optic disc, and it is also noticeable