5 Segmentation of OCT Scans Using Probabilistic Graphical Models
113
(a) 2-D circular scan
(b) 3-D volume
Fig. 5.4 Fundus images that depict a the trajectory and radius of a 2-D circular scan centered around
the optic nerve head and b the area covered by a 3-D volume consisting of 61 fovea-centered BScans. Alternating coloring illustrates the partitioning into 17 different regions. For each region a
separate model is trained
The 3-D datasets consist of fovea-centered volumes. Our in-house dataset contains
35 subjects and each volume is composed of 61 B-Scans with 768 × 496 pixel.
To exclude the nerve head, our model covers a smaller area of 500 × 496 pixel,
corresponding to approximately 5.7 × 7.3 mm. Ground truth was obtained as follows:
Each volume was divided into 17 regions, and a B-scan randomly drawn from each
region was labeled. Figure 5.4b depicts the location of all 61 B-Scans and their
partition into regions indicated by color. Both datasets of Tian et al. consist of 10
volumes each, with 10 and 5 B-Scans labeled. The ground truth provided is a subset
of the surfaces labeled by us. The second dataset consists of subjects with mild
non-proliferative diabetic retinopathy, an early stage of the disease with only small
deformations, which our approach can handle easily.
5.3.1.2 Model Parameters
Table 5.2 summarizes the model parameters and the values they were set to for all
experiments. For the appearance models we set α glasso to 0.01, a parameter of the
glasso approach [19] that controls the sparseness of
−1
x i, j
. A patch-size of 15 × 15 and
the projection onto the first q pca = 20 eigenvectors resulted in smooth segmentation
boundaries. Similar, we used q ppca = 20 eigenvectors to build the shape prior model,
after examining the eigenvalue spectrum of the empirical covariance matrix S.
Table 5.2 Set of model parameter values used throughout all experiments
Parameters
Appearance
Shape
Inference
α glasso
q pca
Patch-Size
q ppca
Variance of
p(b k, j |b \ j )
Value
0.01
20
15 × 15
20
10
113
(a) 2-D circular scan
(b) 3-D volume
Fig. 5.4 Fundus images that depict a the trajectory and radius of a 2-D circular scan centered around
the optic nerve head and b the area covered by a 3-D volume consisting of 61 fovea-centered BScans. Alternating coloring illustrates the partitioning into 17 different regions. For each region a
separate model is trained
The 3-D datasets consist of fovea-centered volumes. Our in-house dataset contains
35 subjects and each volume is composed of 61 B-Scans with 768 × 496 pixel.
To exclude the nerve head, our model covers a smaller area of 500 × 496 pixel,
corresponding to approximately 5.7 × 7.3 mm. Ground truth was obtained as follows:
Each volume was divided into 17 regions, and a B-scan randomly drawn from each
region was labeled. Figure 5.4b depicts the location of all 61 B-Scans and their
partition into regions indicated by color. Both datasets of Tian et al. consist of 10
volumes each, with 10 and 5 B-Scans labeled. The ground truth provided is a subset
of the surfaces labeled by us. The second dataset consists of subjects with mild
non-proliferative diabetic retinopathy, an early stage of the disease with only small
deformations, which our approach can handle easily.
5.3.1.2 Model Parameters
Table 5.2 summarizes the model parameters and the values they were set to for all
experiments. For the appearance models we set α glasso to 0.01, a parameter of the
glasso approach [19] that controls the sparseness of
−1
x i, j
. A patch-size of 15 × 15 and
the projection onto the first q pca = 20 eigenvectors resulted in smooth segmentation
boundaries. Similar, we used q ppca = 20 eigenvectors to build the shape prior model,
after examining the eigenvalue spectrum of the empirical covariance matrix S.
Table 5.2 Set of model parameter values used throughout all experiments
Parameters
Appearance
Shape
Inference
α glasso
q pca
Patch-Size
q ppca
Variance of
p(b k, j |b \ j )
Value
0.01
20
15 × 15
20
10