106
F. Rathke et al.
tion of subsequent boundaries is guided by restricting their search region. They also
demonstrated the applicability of their approach to the segmentation of pathological
scans [9, 10]. Tian et al. [11, 12], which use an adaptation of [8], published two OCT
benchmark datasets which we will use for performance evaluation along with their
results as baselines. Finally, Duan et al. [13] proposed a similar approach, with the
exception that the shortest path is found in the continuous domain.
Another set of approaches, including [14–17], build an undirected graphical model
that includes more than one boundary. Here shape regularization is imposed on the
pairwise interaction of adjacent boundaries, constraining their relative positions. This
form of shape prior information is encoded into the model either as hard constraints
[14] or as probabilistic soft constraints [15, 16]. Inference is performed via graph
cuts. Due to computational limitations, only local shape information is used and not
all boundaries are inferred at the same time.
Finally, Kaji` c et al. [18] apply the popular active appearance approach for matching statistical models of appearance and shape to a given OCT scan. Although nonlocal shape modeling is within the scope of their approach, only landmarks at sparsely
sampled boundary positions are used, and only a maximum likelihood point estimate
is inferred, rather than employing a global probabilistic shape model as does our
approach.
Contribution. We present a probabilistic approach to the segmentation of OCT retina
scans (See Fig. 5.1 for the segmented retina layers). We utilize a global shape prior
that takes into account both short and long-term dependencies between retina layers,
which has not be explored in related work so far. All boundaries are detected at the
same time, which makes obsolete additional processing steps required in other work,
like detection of the fovea, flattening of the retina, or denoising. Since imposing global
shape information is not justified when dealing with pathological deformations, we
outline ongoing work that demonstrates how the model can be adapted to such cases.
Organization. We introduce in Sect. 5.2 our graphical model and explain how inference is done by evaluating the posterior distribution of segmentations. A thorough
evaluation on five different datasets, composed of healthy data and mild pathologies,
follows in Sect. 5.3. We highlight benefits of inference based on a full probability
distribution. In Sect. 5.4, we discuss an extension of our approach that also handles
severe pathological deformations. We conclude and point out further directions of
research in Sect. 5.5.
5.2 A Probabilistic Graphical Model for Retina
Segmentation
5.2.1 The Graphical Model
Mathematically, an OCT scan is a matrix y ∈ R
N ×M of N rows and M columns, containing gray values y i, j in the range [0, 1] = {x ∈ R | 0 ≤ x ≤ 1}. A segmentation
F. Rathke et al.
tion of subsequent boundaries is guided by restricting their search region. They also
demonstrated the applicability of their approach to the segmentation of pathological
scans [9, 10]. Tian et al. [11, 12], which use an adaptation of [8], published two OCT
benchmark datasets which we will use for performance evaluation along with their
results as baselines. Finally, Duan et al. [13] proposed a similar approach, with the
exception that the shortest path is found in the continuous domain.
Another set of approaches, including [14–17], build an undirected graphical model
that includes more than one boundary. Here shape regularization is imposed on the
pairwise interaction of adjacent boundaries, constraining their relative positions. This
form of shape prior information is encoded into the model either as hard constraints
[14] or as probabilistic soft constraints [15, 16]. Inference is performed via graph
cuts. Due to computational limitations, only local shape information is used and not
all boundaries are inferred at the same time.
Finally, Kaji` c et al. [18] apply the popular active appearance approach for matching statistical models of appearance and shape to a given OCT scan. Although nonlocal shape modeling is within the scope of their approach, only landmarks at sparsely
sampled boundary positions are used, and only a maximum likelihood point estimate
is inferred, rather than employing a global probabilistic shape model as does our
approach.
Contribution. We present a probabilistic approach to the segmentation of OCT retina
scans (See Fig. 5.1 for the segmented retina layers). We utilize a global shape prior
that takes into account both short and long-term dependencies between retina layers,
which has not be explored in related work so far. All boundaries are detected at the
same time, which makes obsolete additional processing steps required in other work,
like detection of the fovea, flattening of the retina, or denoising. Since imposing global
shape information is not justified when dealing with pathological deformations, we
outline ongoing work that demonstrates how the model can be adapted to such cases.
Organization. We introduce in Sect. 5.2 our graphical model and explain how inference is done by evaluating the posterior distribution of segmentations. A thorough
evaluation on five different datasets, composed of healthy data and mild pathologies,
follows in Sect. 5.3. We highlight benefits of inference based on a full probability
distribution. In Sect. 5.4, we discuss an extension of our approach that also handles
severe pathological deformations. We conclude and point out further directions of
research in Sect. 5.5.
5.2 A Probabilistic Graphical Model for Retina
Segmentation
5.2.1 The Graphical Model
Mathematically, an OCT scan is a matrix y ∈ R
N ×M of N rows and M columns, containing gray values y i, j in the range [0, 1] = {x ∈ R | 0 ≤ x ≤ 1}. A segmentation