10 Layer Segmentation and Analysis for Retina with Diseases
259
stitial fluid accumulation [37], swelling and thickening of the macular layers, and
finally damage to central vision [38, 39]. The external limiting membrane (ELM)
is a structure that separates the inner segments from the outer nuclear layer, where
the Müller cells are joined to the photoreceptor cells. The ELM serves as a skeleton to keep the photoreceptors aligned [40]. The ELM has been hypothesized to
maintain a protein balance between the photoreceptor layer and the outer nuclear
layer [41]. Recently, several studies have shown that ELM interruptions visible on
spectral-domain optical coherence tomography (SD-OCT) are associated with lower
visual acuity outcome in patients with clinically significant diabetic macular edema
(CSME) [42–45]. Possibly this is because the integrity of the ELM has a critical role
in restoration of the photoreceptor microstructures and alignment [46–48]. Earlier
reported approaches [49–51] relied on manual tracing of the corresponding surface
or on detecting ELM on 2-D B-scans. However, such studies rely on manual interpretation of the state of the ELM, and high intra- and inter-observer variabilities are
likely. Automated 3-D analysis of the ELM is of high interest because of its potential
to elucidate structural abnormalities with minimal variability and possibly predict
visual outcomes in diabetic macular edema (DME).
In this part we describe a novel and fully automated method to quantify the
integrity of the ELM in patients with CSME and in normal subjects based on SDOCT volumes [23]. This pilot study showing differences between 16 normal controls
and 16 CSME patients demonstrates the practical feasibility of the presented methodology and provide preliminary comparisons between these two groups of subjects.
10.3.2 Method
10.3.2.1 ELM Layer Segmentation
The Iowa reference algorithm based on graph search [1, 2, 14, 28, 29] is first applied
to segment the OCT volume, yielding 11 surfaces (Fig. 10.9). Then, the subvolume
between surfaces 6 and 11 (region between OPL and the ONL), which contains the
ELM, was flattened based on the segmented RPE floor (surface 11). Subsequently,
the graph search surface-detection method is applied again to segment the ELM layer
in this subvolume.
10.3.2.2 ELM Disruption Area Detection
In this section, each A-scan is classified as disrupted or nondisrupted based on the
texture and morphology in the vicinity of the ELM surface. The original OCT images
are first enhanced by standard normalization. Six texture features are then extracted
for classification, including intensity, gradient, local variance, local intensity orientation, local coherence, and retinal thickness. The intensity represents the voxel’s
gray-level intensity; the gradient represents the intensity difference between the voxel
259
stitial fluid accumulation [37], swelling and thickening of the macular layers, and
finally damage to central vision [38, 39]. The external limiting membrane (ELM)
is a structure that separates the inner segments from the outer nuclear layer, where
the Müller cells are joined to the photoreceptor cells. The ELM serves as a skeleton to keep the photoreceptors aligned [40]. The ELM has been hypothesized to
maintain a protein balance between the photoreceptor layer and the outer nuclear
layer [41]. Recently, several studies have shown that ELM interruptions visible on
spectral-domain optical coherence tomography (SD-OCT) are associated with lower
visual acuity outcome in patients with clinically significant diabetic macular edema
(CSME) [42–45]. Possibly this is because the integrity of the ELM has a critical role
in restoration of the photoreceptor microstructures and alignment [46–48]. Earlier
reported approaches [49–51] relied on manual tracing of the corresponding surface
or on detecting ELM on 2-D B-scans. However, such studies rely on manual interpretation of the state of the ELM, and high intra- and inter-observer variabilities are
likely. Automated 3-D analysis of the ELM is of high interest because of its potential
to elucidate structural abnormalities with minimal variability and possibly predict
visual outcomes in diabetic macular edema (DME).
In this part we describe a novel and fully automated method to quantify the
integrity of the ELM in patients with CSME and in normal subjects based on SDOCT volumes [23]. This pilot study showing differences between 16 normal controls
and 16 CSME patients demonstrates the practical feasibility of the presented methodology and provide preliminary comparisons between these two groups of subjects.
10.3.2 Method
10.3.2.1 ELM Layer Segmentation
The Iowa reference algorithm based on graph search [1, 2, 14, 28, 29] is first applied
to segment the OCT volume, yielding 11 surfaces (Fig. 10.9). Then, the subvolume
between surfaces 6 and 11 (region between OPL and the ONL), which contains the
ELM, was flattened based on the segmented RPE floor (surface 11). Subsequently,
the graph search surface-detection method is applied again to segment the ELM layer
in this subvolume.
10.3.2.2 ELM Disruption Area Detection
In this section, each A-scan is classified as disrupted or nondisrupted based on the
texture and morphology in the vicinity of the ELM surface. The original OCT images
are first enhanced by standard normalization. Six texture features are then extracted
for classification, including intensity, gradient, local variance, local intensity orientation, local coherence, and retinal thickness. The intensity represents the voxel’s
gray-level intensity; the gradient represents the intensity difference between the voxel