Chapter 5
Segmentation of OCT Scans Using
Probabilistic Graphical Models
Fabian Rathke, Mattia Desana and Christoph Schnörr
The most prominent structure in retinal OCT images is the intra-retinal layers. Therefore layer segmentation is one of the most studied problems in OCT image processing.
In this chapter, a probabilistic approach for retinal layer segmentation is presented.
It exploits texture and shape information based on a graphical model. The approach
can be extended to a locally adaptive graphical model that additionally discriminates
between healthy and pathologically deformed scan structure.
5.1 Introduction
Since its introduction in 1991 [1], optical coherence tomography (OCT) has become
a standard tool in clinical ophthalmology [2]. The introduction of spectral-domain
OCT [3] dramatically increased the resolution as well as the imaging speed and
enabled the acquisition of 3-D volumes composed of hundreds of 2-D scans. Since
the manual segmentation of retina scans is tedious and time-consuming, automated
segmentation methods become evermore important given the growing amount of
gathered data.
Related Work. Various segmentation approaches have been published. All have in
common that they utilize texture information, based on the spatial variation of the
intensity functions and its gradient. In order to obtain accurate and stable segmentations and to reduce the sensitive to texture artifacts, some form of regularization
is applied. We focus on the methods used for regularization in related work and to
which degree shape prior knowledge is utilized.
Many approaches solely impose smoothness on segmented retina boundaries,
without any shape prior information. References [4–7] find column-wise maxima
of the appearance terms and then apply outlier detection along with interpolation
to account for erroneous segmentations. Chiu et al. [8] construct a graph for each
boundary with weights determined by gradient information and find the shortest path
using dynamic programming. Starting with easy to detect boundaries, the segmentaF. Rathke (B) · M. Desana · C. Schnörr
Image and Pattern Analysis Group (IPA), University of Heidelberg, Heidelberg, Germany
e-mail: fabian.rathke@iwr.uni-heidelberg.de
© Science Press and Springer Nature Singapore Pte Ltd. 2019
X. Chen et al. (eds.), Retinal Optical Coherence Tomography Image Analysis,
Biological and Medical Physics, Biomedical Engineering,
https://doi.org/10.1007/978-981-13-1825-2_5
105
Segmentation of OCT Scans Using
Probabilistic Graphical Models
Fabian Rathke, Mattia Desana and Christoph Schnörr
The most prominent structure in retinal OCT images is the intra-retinal layers. Therefore layer segmentation is one of the most studied problems in OCT image processing.
In this chapter, a probabilistic approach for retinal layer segmentation is presented.
It exploits texture and shape information based on a graphical model. The approach
can be extended to a locally adaptive graphical model that additionally discriminates
between healthy and pathologically deformed scan structure.
5.1 Introduction
Since its introduction in 1991 [1], optical coherence tomography (OCT) has become
a standard tool in clinical ophthalmology [2]. The introduction of spectral-domain
OCT [3] dramatically increased the resolution as well as the imaging speed and
enabled the acquisition of 3-D volumes composed of hundreds of 2-D scans. Since
the manual segmentation of retina scans is tedious and time-consuming, automated
segmentation methods become evermore important given the growing amount of
gathered data.
Related Work. Various segmentation approaches have been published. All have in
common that they utilize texture information, based on the spatial variation of the
intensity functions and its gradient. In order to obtain accurate and stable segmentations and to reduce the sensitive to texture artifacts, some form of regularization
is applied. We focus on the methods used for regularization in related work and to
which degree shape prior knowledge is utilized.
Many approaches solely impose smoothness on segmented retina boundaries,
without any shape prior information. References [4–7] find column-wise maxima
of the appearance terms and then apply outlier detection along with interpolation
to account for erroneous segmentations. Chiu et al. [8] construct a graph for each
boundary with weights determined by gradient information and find the shortest path
using dynamic programming. Starting with easy to detect boundaries, the segmentaF. Rathke (B) · M. Desana · C. Schnörr
Image and Pattern Analysis Group (IPA), University of Heidelberg, Heidelberg, Germany
e-mail: fabian.rathke@iwr.uni-heidelberg.de
© Science Press and Springer Nature Singapore Pte Ltd. 2019
X. Chen et al. (eds.), Retinal Optical Coherence Tomography Image Analysis,
Biological and Medical Physics, Biomedical Engineering,
https://doi.org/10.1007/978-981-13-1825-2_5
105