AbstractOptical coherence tomography (OCT) is an
important mode of biomedical imaging for the diagnosis and
management of ocular disease. Here we report on the
construction of a synthetic retinal OCT image data set that may
be used for quantitative analysis of image processing methods.
Synthetic image data were generated from statistical
characteristics of real images (n = 14). Features include:
multiple stratified layers with representative thickness,
boundary gradients, contour, and intensity distributions
derived from real data. The synthetic data also include retinal
vasculature with typical signal obscuration beneath vessels.
This synthetic retinal image can provide a realistic simulated
data set to help quantify the performance of image processing
algorithms.
I. INTRODUCTION
PTICAL coherence tomography (OCT) imaging of the
eye has rapidly become one of the most important
imaging modalities for the diagnosis and management of
retinal diseases. Current state of the art Spectral Domain
OCT (SD-OCT) imaging systems provide clinicians with
micron-level resolution [1-3]. This technology has moved
rapidly from engineering to clinical implementation and
tools to quantify and assist users with visualization and
interpretation of the data generated are still developing [4-7].
Quantitative analysis of biomedical image data plays an
increasingly important role in medical decision making.
Feature detection may be an important component of tumor
identification in breast cancer [8-13] and image
segmentation may be critical information that helps
clinicians to decide whether or not medical therapy is halting
the progression of optic nerve degeneration in glaucoma [14,
15]. Biomedical image information is also increasingly used
as guidance for surgical treatments [16]. Each of these
applications relies on various elements of image processing
that may include feature detection, segmentation,
registration, and denoising as well as other operations.
Reliable quantitative analysis of image data is an important
step in deriving actionable information from imaging data.
Manuscript received April 15, 2011.
P. M. Kulkarni, Electrical and Computer Engineering, University of
Houston, Houston, TX 77204 USA. (e-mail:
prathameshmkulkarni@gmail.com).
D. C. Lozano, University of Houston College of Optometry, Houston,
TX 77204-2020 USA (e-mail: lozano.dc@gmail.com).
G. Zouridakis, Engineering Technology, Computer Science, and
Electrical and Computer Engineering, University of Houston, Houston, TX
77204 USA (e-mail: zouridakis@uh.edu).
M. Twa, University of Houston College of Optometry, Houston, TX
77204-2020 USA (713-743-2996; e-mail: mdtwa@uh.edu).
There are numerous studies that use synthetic or phantom
images as a benchmark to enable quantitative comparisons
of image processing operations. For example, the modified
Shepp-Logan phantom is frequently used as a reference for
CT and MRI image processing evaluations. Despite its
simplicity (or perhaps because of it) this model has been a
highly useful tool for quantitative analysis of image
processing methods in MRI and CT data [17, 18]. Phantom
images have also been used for validation of image
registration [20] and segmentation algorithms [21, 22] for
ultrasound data. However, these synthetic images were not
derived from real data. In this research, we derive a synthetic
model for retinal OCT data from real image data. The
objective of this research was to develop a model of retinal
SD-OCT data that could be used as a benchmark for similar
quantitative research with OCT image data. The goal was to
provide a standardized synthetic data set for comparison of
various image processing methods.
II. METHODS
A. Data
Fourteen high-quality B-Scan images were selected from
within three different C-Scan volumes, each from a different
animal. All parameters of the model were derived using
these fourteen images. To insure that all B-Scan images had
comparable thickness parameters for each layer, the B-Scan
images were selected from regions adjacent to the optic
nerve head (Figure 1).
B. Manual segmentation
Each B-Scan image included in the model was manually
segmented by domain experts using a custom designed
image annotation tool. The segmentation procedure involved
delineating the boundaries for seven different retinal layers
and demarcating blood vessels in the image. The seven
layers included in the segmentations were: Nerve Fiber
Layer-Ganglion Cell Layer (NFL-GCL), Inner Plexiform
Layer (IPL), Inner Nuclear Layer (INL), Outer Plexiform
Layer (OPL), Outer Nuclear Layer (ONL), Photo Receptor
Layer (PRL) and Retinal Pigmented Epithelium (RPE). Each
of these layers is identified separately in Figure 2. Manual
segmentation of these multiple images was the basis for the
intensity, gradient and contour parameters of the synthetic
image model.
C. Intensity modeling
Intensity distribution models were derived for seven retinal
layers by analyzing the distributions for these layers across
A Statistical Model of Retinal Optical Coherence Tomography
Image Data
Prathamesh Kulkarni, Diana Lozano, George Zouridakis, and Michael Twa
O
978-1-4244-4122-8/11/$26.00 ©2011 IEEE 6127
33rd Annual International Conference of the IEEE EMBS
Boston, Massachusetts USA, August 30 - September 3, 2011