11 Segmentation and Visualization of Drusen …
307
11.3.1.3 Generation of GA Projection Image
A common method for creating a 2D projection from SD-OCT volumetric datasets
is the summed-voxel projection (SVP) [36], in which all the voxel values in the 3D
data are summed along the axial A-scan lines in the B-scans, producing an image
showing the retinal surface en face, similar to color fundus photographs (CFPs) and
FAF images. However, the en face SVP fundus image is not ideal for GA visualization
due to the confounding influence of highly reflective retinal layers above and below
GA lesions in the retina (in particular the RNFL and RPE layers) which obscure GA
lesions. The commercial software on the Cirrus HD-OCT (version 6.0) provides a
sub-RPE slab function [62]. The sub-RPE slab is formed by axially projecting only
the OCT image data from a region below the contour of the RPE fit. Our projection
method, which is also derived by restricting the sum of the voxel values to the subvolume beneath the segmented RPE layer, where the choroid resides and where the
high reflections indicating GA will be seen, improves the traditional SVP image
in terms of GA visualization. The lower boundary of the sub-volume is parallel to
the top boundary (RPE layer), where the parallel distance is equal to the minimum
distance between the end of the cube and the segmented RPE layer. The average
intensity of the sub-volume in the axial direction is taken as the intensity value of the
GA projection image. We call this the restricted summed-voxel projection (RSVP).
Figure 11.16 shows an example, comparing the traditional SVP projection [36]
and the RSVP projection in a patient with GA. Figure 11.3 shows that the contrast
of GA in the RSVP image is higher than in the SVP image, which can improve the
performance of a computerized GA segmentation method. On the other hand, this
process can introduce aberrant bright signals (e.g., the bright spots near the upper
blood vessels in Fig. 11.16), caused by an inaccurate RPE layer segmentation. In
practice, this did not negatively impact our results (see the evaluation of our GA
segmentation method below).
11.3.1.4 GA Segmentation Based on Geometric Active Contour Model
For derivation of the shape and size of GA lesions, we used geometric active contour
model for the segmentation of the GA lesions on the RSVP images. These images
were denoised using bilateral filtering [45] as a preliminary step to reduce the influence of noise on the segmentations. Geometric active contour (GAC) models were
simultaneously proposed by Caselles [63] and by Malladi [64], introduced as an
alternative to parametric deformable models and as a way to overcome their limitations. GAC models are based on the theory of curve evolution and geometric flows,
and implemented using the level-sets based numerical algorithm. The basic idea of
these models is to transform a planar curve movement track into a three-dimensional
curved surface movement track, which has the advantage of being able to handle the
change of topological structure easily. During the evolution of traditional level set
methods, re-initialization is necessary to keep the evolving level set function close to
a signed distance function. In order to eliminate the need of the costly re-initialization
307
11.3.1.3 Generation of GA Projection Image
A common method for creating a 2D projection from SD-OCT volumetric datasets
is the summed-voxel projection (SVP) [36], in which all the voxel values in the 3D
data are summed along the axial A-scan lines in the B-scans, producing an image
showing the retinal surface en face, similar to color fundus photographs (CFPs) and
FAF images. However, the en face SVP fundus image is not ideal for GA visualization
due to the confounding influence of highly reflective retinal layers above and below
GA lesions in the retina (in particular the RNFL and RPE layers) which obscure GA
lesions. The commercial software on the Cirrus HD-OCT (version 6.0) provides a
sub-RPE slab function [62]. The sub-RPE slab is formed by axially projecting only
the OCT image data from a region below the contour of the RPE fit. Our projection
method, which is also derived by restricting the sum of the voxel values to the subvolume beneath the segmented RPE layer, where the choroid resides and where the
high reflections indicating GA will be seen, improves the traditional SVP image
in terms of GA visualization. The lower boundary of the sub-volume is parallel to
the top boundary (RPE layer), where the parallel distance is equal to the minimum
distance between the end of the cube and the segmented RPE layer. The average
intensity of the sub-volume in the axial direction is taken as the intensity value of the
GA projection image. We call this the restricted summed-voxel projection (RSVP).
Figure 11.16 shows an example, comparing the traditional SVP projection [36]
and the RSVP projection in a patient with GA. Figure 11.3 shows that the contrast
of GA in the RSVP image is higher than in the SVP image, which can improve the
performance of a computerized GA segmentation method. On the other hand, this
process can introduce aberrant bright signals (e.g., the bright spots near the upper
blood vessels in Fig. 11.16), caused by an inaccurate RPE layer segmentation. In
practice, this did not negatively impact our results (see the evaluation of our GA
segmentation method below).
11.3.1.4 GA Segmentation Based on Geometric Active Contour Model
For derivation of the shape and size of GA lesions, we used geometric active contour
model for the segmentation of the GA lesions on the RSVP images. These images
were denoised using bilateral filtering [45] as a preliminary step to reduce the influence of noise on the segmentations. Geometric active contour (GAC) models were
simultaneously proposed by Caselles [63] and by Malladi [64], introduced as an
alternative to parametric deformable models and as a way to overcome their limitations. GAC models are based on the theory of curve evolution and geometric flows,
and implemented using the level-sets based numerical algorithm. The basic idea of
these models is to transform a planar curve movement track into a three-dimensional
curved surface movement track, which has the advantage of being able to handle the
change of topological structure easily. During the evolution of traditional level set
methods, re-initialization is necessary to keep the evolving level set function close to
a signed distance function. In order to eliminate the need of the costly re-initialization