Choroid Characterization in EDI OCT Retinal

Images Based on Texture Analysis

A. Gonz

alez-L

opez, B. Remeseiro, M. Ortega and M. G. Penedo

Departmento de Computaci

on, Universidade da Coru

na, A Coru

na 15071, Spain

Keywords:

OCT, Retinal Images, Choroid, Texture Analysis, Pattern Recognition, Machine Learning.

Abstract:

Optical Coherence Tomography (OCT) is a widely extended imaging technique in the opthalmic ﬁeld for diag-

nostic purposes. Since layers composing retina can be identiﬁed in these images, several image processing-

based methods have been presented to segment them automatically in these images, with the aim of developing

medical-support applications. Recently, appearance of Enhanced Depth Imaging (EDI) OCT allows to tackle

exploration of the choroid which provides high information of eye processes. Therefore, segmentation of

choroid layer has become one of the more relevant problems tackled in this ﬁeld, but it presents different fea-

tures that rest of the layers. In this work, a novel texture-based study is proposed in order to show that textural

information can be used to characterize this layer. A pattern recognition process is carried out by using diffe-

rent descriptors and a process of classiﬁcation, considering marks performed by two experts for validation.

Results show that characterization using texture features is effective with rates over 90% of success.

1 INTRODUCTION

Optical Coherence Tomography (OCT) is a standard

imaging technique in the ophthalmologic ﬁeld, which

provides a cross sectional image of the retina in a

non-invasive, real time fashion (Puzyeyeva et al.,

2011). Experts use OCT retinal images to diagnose

diseases, because retinal morphology can be identi-

ﬁed effectively on them, explaining disease pathoge-

nesis and progression. Nowadays, Enhanced Depth

Imaging (EDI) OCT allows to better visualize the

choroid, which is the vascular tissue located at the

posterior part of the eye, between the retina and the

sclera. Choroid provides oxygen and nourishment to

the outer layers of the retina (Bill et al., 1983) and

its changes have been hypothesized to be of critical

importance in the pathophysiology of several retinal

diseases, such as glaucoma (Yin et al., 1997), cen-

tral serous chorioretinopathy (Imamura et al., 2009)

or retinitis pigmentosa (Dhoot et al., 2013). There-

fore, characterization and delimitation of this layer in

the OCT retinal images is important to understand the

natural processes of the eye, besides detecting poten-

tial eye diseases.

Although retinal segmentation is a widely studied

problem, automatical choroidal layer detection has

not been investigated in as much depth. A few me-

thods for this task have been recently reported. In

(Kaji

c et al., 2013), a two stage elaborated statis-

tical model is presented to automatically detect the

choroidal boundaries in EDI OCT images. A two-

stage active contour-based technique is used in (Lu

et al., 2013) to segment the outer boundary of the

choroid, but it requires the manual initialization of

the contour, while another semiautomatic approach is

also studied in (Hu et al., 2013). In (Alonso-Caneiro

et al., 2013), a graph-based method obtains promis-

ing results, but a process of enhancement is conside-

red, dependant of the instrument used for the image

capture. Choroidal vessels were analyzed through

a choroidal segmentation method in (Zhang et al.,

2012), but the focus of this work was to quantify the

vasculature rather than the choroidal thickness.

Since choroidal surface presents different visual

properties than the rest of retinal layers, a texture-

based characterization is interesting to be tackled, as

a previous step for a future segmentation method for

choroidal boundaries. Although texture information

is included in previous work (Danesh et al., 2014),

where a Gaussian Mixture Model of the image is built

based on features extracted with the Discrete Wavelet

Transform, and then used in the segmentation task, it

has not been studied deeply which features are those

that best describes this layer for a future and robust

process of segmentation. With that purpose, this work

presents a study of texture features that can be consi-

269

Gonzalez-Lopez A., Remeseiro B., Ortega M. and Penedo M..

Choroid Characterization in EDI OCT Retinal Images Based on Texture Analysis.

DOI: 10.5220/0005177602690276

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 269-276

ISBN: 978-989-758-074-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

dered to characterize this layer, in order to determine

those that must be used to represent information in

this layer and design future process of segmentation

and information extraction. As results show, choroid

can be characterized effectively using textural infor-

mation.

This paper is organized as follows: Section 2 ex-

plains the method designed for the characterization of

the choroid. Section 3 describes materials and me-

thods, while in Section 4 obtained results are presen-

ted. Finally, conclusions and future lines are presen-

ted in Section 5.

2 METHODOLOGY

As it was introduced in Section 1, this work aims to

characterize choroidal layer using texture features, in

order to allow a future segmentation process of this

layer.

The process is described as follows: ﬁrstly,

the image provided by the OCT scanner is prepro-

cessed, in order to make it suitable for the method.

Since choroid is located between the Retinal Pig-

ment Epithelium-Bruch’s Membrane complex (RPE-

BM) and the sclera (see Figure 1), the region of in-

terest (ROI) must be bounded. Thus, choroidal up-

per boundary is detected and only the area covering

choroid and sclera is studied. After that, several win-

dows are extracted from the image to build diffe-

rent texture descriptors. Since ophthalmologic ex-

perts have marked manually the boundary between

the choroid and the sclera (marked in Figure 1 as

Outer Choroid Boundary (OCB) ) over the images,

extracted features can be used to perform a classiﬁ-

cation process, in order to characterize the choroidal

layer. Thus, it is possible to perform a pattern recog-

nition process to discriminate between regions be-

longing to the choroid and those located in the sclera.

This process is reﬂected graphically in Figure 2.

Choroid

RPE-BM

Sclera

OCB

Figure 1: Sample OCT image with choroidal layer bounda-

ries marked in red.

2.1 Preprocessing

Images provided by the OCT device includes irrele-

vant areas which must be excluded. Besides that, they

are captured following an inverted gray-scale distribu-

tion than that used in the segmentation process needed

in the ROI bounding step (Section 2.2). Thus, they

are automatically cropped and inverted to make the

following step feasible. It is essential to consider that

image captured by the scanner is resized for a better

visualization. This involves that, as it was observed

during initial test, decreasing resolution of the images

provided better results when texture features are ex-

tracted. Therefore, to tackle this process, after crop-

ping images, they are resized using a scale factor of

0.5. Consequently, not only the effectiveness of the

process is increased, but also the computation time is

reduced.

2.2 ROI Location

As it was commented, ROI is determined by the area

covering choroid and sclera. In order to bound it, the

inner boundary of the choroid surface, corresponding

to its borderline with the RPE-BC (see Figure 1), must

be detected. With that purpose, the multistage ac-

tive contour-based segmentation technique described

in (Gonz

alez et al., 2014) is used. This method detects

the boundary between these layers through a process

of energy minimization.

In particular, the active contour model used to seg-

ment this boundary is described as follows:

• Topology: it is deﬁned as a sequence of nodes co-

vering the image width. Each node corresponds to

one pixel in the image and has two neighbors, ex-

cept for the ﬁrst and the last ones. During the pro-

cess of minimization, nodes can make displace-

ments to its 8-connected neighbors (except for

nodes in the extrema, which only can move along

the rows in the image).

• Internal energy: ﬁrst and second order terms are

considered in order to guarantee continuity and

curvature.

• External energy: since the boundary of interest

can be identiﬁed as a light-to-dark transition in the

image, besides the fact that it bounds a wide bright

area (RPE-BC), it includes information of edges

and regional intensity, as text belows explains.

The gradient distance is used, being computed

over edges corresponding to the mentioned kind

of transition. In order to avoid the inﬂuence of

some edges that would make the model reaching

wrong solutions and also for a fast evolution of the

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

270

Figure 2: Phases followed for choroid characterization over a sample EDI OCT image.

contour, a new energy term is added, represent-

ing the distance between a node and the strongest

edge in the area of searching. This term is impor-

tant to encourage the movement of the contour in

the ﬁrst steps of its evolution.

Regarding the intensity-based information, a term

encouraging pixels with bright regions above is

deﬁned: accumulated intensity in a range of rows

above each pixel in the image is computed; then,

those pixels with the highest value per column are

extracted and ﬁnally, the value of this term is com-

puted as the gradient distance over them. These

steps are reﬂected in Figure 3 (f) (g) and (h), res-

pectively, for an easier understanding.

It is necessary to clarify that, in order to combine

different levels of information detail during the pro-

cess, this model is designed to work over two images

at the same time, each one enhanced in a different

manner. The ﬁrst image is the original one smoothed

and with enhanced contrast, which contains precise

information, while the second image, obtained with

an aggressive preprocessing, provides a coarse level

of information, useful in the ﬁrsts steps of energy min-

imization.

Using this model, the nodes in the active contour

will be attracted to the desired boundary in the image

through a process of several stages of minimization.

For this boundary, two stages of minimization have

been deﬁned, allowing to adjust parameters of the

model.

However, given that this method was designed for

standard OCT images, it needs to be modiﬁed in or-

der to make it feasible for EDI-OCT images. The

main difference between images captured with both

techniques is that lower layers present higher con-

trast and deﬁnition in case of EDI OCT. Thus, as pre-

vious step for RPE-BC/Choroid boundary segmen-

tation, the mentioned model is used to detect the

upper boundary of RPE-BC (it corresponds to the

strongest edge in the image). Using that information,

the active contour used to segment RPE-BC/Choroid

is initialized automatically and the segmentation of

the boundary is immediate, considering that it corre-

sponds to a very strong light-to-dark transition. Ob-

viously, parameters of the active contour-based model

must be adjusted in order to extend it for this new kind

of images.

2.3 Choroid Characterization

Once the upper boundary of the choroid is deter-

mined, it is necessary to establish its limit with the

sclera. Choroid layer presents different features than

the rest of layers in the retina. For instance, since

choroid is deep underneath the retina, OCT signals

can be degraded. This fact, in addition with the weak-

ness presented by choroid lower boundary, makes it

almost invisible in most cases. Besides that, its ap-

pearance is different from the other layers, because

it is formed by a dense vascularity structure and its

shape presents greater variations in thickness. In con-

traposition to that, sclera presents a more homoge-

neous intensity and it does not present vascularity fea-

tures.

Considering all these properties, it seems that the

detection of this layer can be tackled as a pattern

recognition problem, where its texture features can

be analyzed to characterize it. With that purpose, af-

ter enhancing the image contrast, several windows of

w × h pixels are obtained from the bounded region in

the images and are used to extract the different tex-

ture descriptors, described in Section 2.3.1 (Figure

4 shows a sample image with some of the extracted

windows). Then, the process of classiﬁcation is per-

formed.

After assessing the performance of the different

texture descriptors at the current problem, the method

providing best results can be explored more deeply.

To do that, a feature selection procedure is applied

with the purpose of determining if removing irrele-

vant or redundant features can involve an improve-

ment in the performance of the classiﬁer.

ChoroidCharacterizationinEDIOCTRetinalImagesBasedonTextureAnalysis

271

(d)

(e)

(b)

(g)

(a)

(f)

(c)

(h)

(i)

Figure 3: Segmentation of the choroid upper boundary

(RPE-BC/Choroid): (a) sample image with initialized ac-

tive contour; some energy terms: (b) edges corresponding

to dark to light transitions; (c) gradient distance computed

over (b); (d) strongest edges computed per column; (e) gra-

dient distance computed over (d); (f) accumulated intensi-

ties above each pixel; (g) strongest pixels per column in (d);

(h) gradient distance computed over (e); (i) boundary given

by interpolation of ﬁnal nodes in the active contour.

(a)

(b)

Figure 4: Sample image with extracted windows correspon-

ding to (a) choroid layer and (b) sclera. Zoom has been

applied in order to show visual textural differences between

both areas (note the different homogeneity presented in both

cases).

2.3.1 Texture Descriptors

In order to analyze different modalities for texture

descriptor extraction, three widely used methods are

studied to perform the analysis, each one correspon-

ding to a different modality. Firstly, Markov Random

Fields (MRF) is analyzed, that is a model-based me-

thod; then, Co-ocurrences Features is proposed as sta-

tistical method and ﬁnally, as signal processing me-

thod, the Discrete Wavelet Transform (DWT). This

last method is chosen also because it is used in pre-

vious work (Danesh et al., 2014). Therefore, this

study is also useful to determine if the approach cho-

sen in that work is the most appropriate for this task

or, on the contrary, there are other methods that can

characterize better the choroid layer.

The three considered methods are detailed as

follows:

Markov Random Fields (MRF). (Besag, 1974) are

model based texture analysis methods that construct

an image model whose parameters capture the

essential perceived qualities of texture. A MRF is a

2D lattice of points where each point is assigned a

value that depends on its neighboring values. Thus,

MRFs generate a texture model by expressing the

gray values of each pixel in an image as a function of

the gray values in a neighborhood of the pixel. The

neighborhood of a pixel is deﬁned as the set of pixels

within a Chebyshev distance d. Once the parameters

of the model are calculated, the descriptor of an

input image is obtained by computing the directional

variances proposed in (C¸ esmeli and Wang, 2001).

Notice that for a distance d, the descriptor comprises

4d features.

Co-occurrence Features. (Haralick et al., 1973) are

based on the computation of the conditional joint

probabilities of all pairwise combinations of gray

levels, given an interpixel distance and an orienta-

tion. This method generates a set of Gray Level

Co-occurrence Matrices, and extract several statistics

from their elements. As in the above method, the

Chebyshev distance is considered and so, for a dis-

tance d = 1, 4 orientations are considered (0

, 45

and 135

, as diagram in Figure 5 reﬂects), and

4 matrices are generated. In general, the number of

orientations and, accordingly, the number of matrices

for a distance d is 4d. From each co-occurrence ma-

trix a set of 14 statistics are computed, representing

features such as homogeneity or contrast. Then, mean

and range across matrices are extracted, resulting in a

set of 28 features which will be the descriptor of the

input image.

The Discrete Wavelet Transform. (Mallat, 1989)

generates a set of wavelets by scaling and translating

a mother wavelet, which is a function deﬁned both

in the spatial and frequency domain. The different

parameters of the mother wavelet control the band-

pass of the ﬁlter in order to generate high-pass (H) of

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

272

Figure 5: Orientations for distance d = 1 in Co-occurrence

Features.

low-pass (L) ﬁlters. The wavelet decomposition of an

image consists in applying wavelets horizontally and

vertically to generate 4 subimages at each scale (LL,

LH, HL, HH), which are then subsampled by a factor

of 2. After the decomposition, the process is repeated

n−1 times over the LL subimage, where n is the num-

ber of scales of the method. This iterative process re-

sults in the so-called standard pyramidal wavelet de-

composition shown in Figure 6.

Among the different mother wavelets found in the

literature, two of the most popular ones have been

used (Daubechies, 1992): Haar is the simplest non-

trivial wavelet, and Daubechies is one representative

type of basis for wavelets. The descriptor of an input

image is obtained by computing the mean and the ab-

solute average deviation of the input and LL images,

and the energy of the LH, HL and HH images. There-

fore, the descriptor of an input image is composed of

2 + 5 × n features.

Figure 6: Pyramidal wavelet decomposition for n=2 scales.

2.3.2 Classiﬁcation

After descriptor extraction, they are processed by a

classiﬁer, which will determine if a region must be-

long to choroid layer or not. In this work, a su-

pervised classiﬁer, Support Vector Machine (SVM),

has been used. SVM is based on the statistical lear-

ning theory, performing classiﬁcation by constructing

an N-dimensional hyperplane that optimally separates

the data in categories. Results in terms of percentage

accuracy are shown in next section, representing the

rate per cent of the windows correctly classiﬁed, ac-

cording to their category (in this case, belonging to

choroid layer or not).

2.3.3 Feature Selection

Once all textural descriptors have been assessed, a

feature selection process is applied over the method

providing the best performance. This is motivated by

the fact that an appropriate selection of the considered

features can lead to an improvement in the classiﬁca-

tion, not only in terms of speed, but also regarding

the generalization capacity or simplicity of the model

(Bol

on-Canedo et al., 2013). In this work, despite

the variety of existing methods for feature selection,

ﬁlters are chosen, because they are computationally

simple and fast. In particular, three ﬁlters are consi-

dered:

• INTERACT (Zhao and Liu, 2009) is a subset ﬁl-

ter mainly based on symmetrical uncertainty and

the consistency contribution, which indicates how

signiﬁcantly the elimination of a feature can affect

consistency.

• Correlation-based Feature Selection (Hall, 1999)

is a simple multivariate ﬁlter that ranks feature

subsets according to a heuristic based on corre-

lation. Thus, subsets containing features that are

highly correlated with the class and uncorrelated

with each other, are selected.

• Consistency-based Filter (Dash and Liu, 2003) is

based on the evaluation of the worth of a subset

of features by the level of consistency in the class

values when the samples are projected onto the

subset of attributes. The inconsistency criterion

allows to determine what extent the dimensionally

reduced data can be accepted.

3 MATERIAL AND METHODS

The aim of this work is to characterize choroid layer

in the OCT retinal images through a texture analysis.

The materials and methods used in this research are

presented in the text that follows.

3.1 Data Source

Experiments have been done over a dataset of 63 two-

dimensional EDI OCT retinal images corresponding

to 7 different patients. Images were provided by Hos-

pital do Barbanza, Ribeira (Spain) and they were ex-

tracted using Spectralis OCT scanner (Heidelberg En-

gineering). Scanner optic axial resolution is 5µm.

Captured images have resolution of 1520 × 496 pi-

xels, although after the processes of cropping and re-

sizing applied during the preprocessing phase, their

dimensions are reduced to 495 × 218 pixels in mean.

ChoroidCharacterizationinEDIOCTRetinalImagesBasedonTextureAnalysis

273

3.2 Experimental Procedure

The experimental procedure is detailed as follows:

1. Extract windows from the ROI in the images,

using marks made by the experts to label them.

2. Apply the three texture analysis methods descri-

bed in Section 2.3.1 to the dataset of images.

3. Train a SVM (Section 2.3.2) with radial basis ker-

nel and automatic parameter estimation. Since 10-

fold cross-validation was used, the average error

across all trials was computed.

4. Evaluate the effectiveness of the proposed me-

thodology in terms of accuracy.

5. Apply the feature-selection procedures over the

textural method providing the best performance.

Repeat steps (3) and (4).

4 RESULTS AND DISCUSSION

Two experts have marked the OCB boundary in all

images. Windows for descriptor extraction have been

obtained and categorized from those areas correspon-

ding to agreement between experts. Window size is

w × h = 31 × 31 pixels, which seems big enough to

cover textural information in the choroid layer. Since

the number of windows extracted from the sclera

area is too much high than those extracted from the

choroid in each image, this set is reduced in order

to consider balanced sets. Regarding the texture des-

criptors used in this work, distances covering range 1

to 7 are studied for Markov Random Fields and Co-

occurrence Features, whereas for DWT, 4 scales and

4 different wavelets are used. Results are shown in

Tables 1, 2 and 3. Best rate for each method has been

highlighted.

Table 1: Accuracy(%) using SVM classiﬁers for Markov

Random Fields with distances d from 1 to 7.

1 2 3 4 5 6 7

80.52 79.59 71.99 77.48 77.38 74.67 75.35

Table 1 presents results for Markov Random

Fields, whose best rate is obtained for distance d =

1 (over 80%). Besides that, it is possible to ob-

serve that bigger the distance, the lower the accuracy,

what means that textural information is progressively

smaller in those cases.

Co-occurrence Features (Table 2) provides rates

over 90% in all cases. In particular, the highest ac-

curacy is near 94% and it is obtained using distance

d = 3.

Table 2: Accuracy(%) using SVM classiﬁer for Co-

ocurrences Features for distances d from 1 to 7

1 2 3 4 5 6 7

93.88 93.80 93.97 93.76 93.62 93.26 93.32

Table 3: Accuracy(%) using SVM classiﬁer for Discrete

Wavelet Transform, considering different mother wavelets

with scales n from 1 to 4. Daubi represents the Daubechies

orthonormal wavelet, with number of vanishing moments

equal to half the coefﬁcient i. Note that the Haar wavelet is

equivalent to Daub2.

1 2 3 4

Haar 86.86 88.68 82.95 80.05

Daub4 86.11 84.02 82.58 79.42

Daub6 85.68 84.31 81.73 78.39

Daub8 85.66 85.86 82.08 78.82

Regarding DWT, Table 3 shows that using Haar

wavelet with n = 2 scales provides the best result

(88.68%). Results also reﬂect that performance de-

creases at the time that distance increases. This is due

to the window size chosen for this work, given that

when more scales are used, information in the small-

est LL subimages is degraded. Therefore, studying

more scales does not make any sense.

Though all considered texture analysis me-

thods present acceptable rates, the highest accuracy

(93.97%) is obtained using Co-occurrence Features

with distance d = 3. Besides that, this method pro-

vides, in general, better and more stable results. Re-

garding Markov Random Fields, most results are

worse than those provided by the other considered

methods. Despite the fact that both Markov Random

Fields and Co-occurrence Features uses information

of the pixel’s neighborhood, it is possible to conclude

that MRF does not work so well because the statis-

tics proposed by Haralick et al. provide much more

information. With regards to DWT, which was spe-

cially interesting because Haar wavelet was used in

the mentioned previous work (Danesh et al., 2014),

it presents lower accuracy (always below 90%) than

Co-ocurrences in all cases.

Table 4: Results using SVM classiﬁer for Co-occurrence

Features with distance d = 3, after applying ﬁlters for fea-

ture selection: ﬁrst row shows number of selected features

whereas the accuracy is shown in second row.

No FS INT CFS Cons

N. Features 28 14 9 15

Accuracy (%) 93.97 93.18 91.35 93.84

Since the method providing the best results, Co-

occurrence Features, considers features that may be

strongly correlated (Haralick et al., 1973), the feature-

selection process proposed in Section 2.3.3 is moti-

vated. Thus, ﬁlters described in that section are consi-

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

274

dered in order to remove irrelevant features. They are

applied to the features extracted using Co-occurrence

Features with distance d = 3, since this method, as

it was commented, provides the best performance.

After feature selection, process described in Section

2.3.2 is repeated: SVM is trained and a 10-fold cross-

validation is done. Table 4 shows accuracy obtained

by each method (INTERACT - INT, correlation-based

feature selection - CFS and consistency-based ﬁlter -

Cons), as well as the number of features selected by

each one.

As it was commented previously, Co-occurrence

Features with distance d = 3 provides the best rate

(93.97%). For this descriptor, 28 features are consi-

dered. Results after applying feature selection, pre-

sented in Table 4, show that CFS has the most aggres-

sive behaviour since it retains 32.14% of the features,

but at the expense of a slight decrease in the accuracy

of the classiﬁer, while Cons and INTERACT retain

around the 50% of the initial features with no degra-

dation in performance. Thus, results obtained with

Co-occurrence Features (d = 3) can be remained but

with a signiﬁcant reduction in the number of features.

Note that a reduction in the number of feature implies

a decrease in the time needed to compute the descrip-

tor. This time, which is not a matter of study in this

paper, could become a bottle neck when applying a

segmentation process in which a great amount of win-

dows have to be analyzed, and herein also lies the re-

levance of using feature selection.

5 CONCLUSIONS AND FUTURE

WORK

In this work, an automatic characterization of the

choroid layer in EDI OCT retinal images based on

textural information has been presented. The me-

thod consists in locating the region of interest through

an active contour-based segmentation and then, ex-

tracting texture descriptors to perform a classiﬁcation

process with the purpose of discriminating between

regions belonging to the choroid from those corres-

ponding to other areas. Different texture descriptors

have been considered for the pattern recognition task

in order to perform a comparative study and deter-

mine the best one to represent this layer. Results show

that the choroid can be identiﬁed in a very effective

way using Co-ocurrences Features, surpassing pre-

vious works (Danesh et al., 2014). Feature selec-

tion has been applied, allowing to eliminate the irre-

levant/redundant features with no degradation in the

accuracy of the performance. Therefore, texture in-

formation extracted with this method can be used to

characterize this layer in a robust way.

As future work, a deeper study of textural infor-

mation with different descriptors could be done, as

well as more classiﬁers can be considered. Besides

that, this information must be included in a methodo-

logy that allows to tackle the accurate segmentation

of the lower choroid boundary.

ACKNOWLEDGEMENTS

This paper has been partly funded by Instituto de

Salud Carlos III under PI14/02161 project and the

Secretar

ıa de Estado de Investigaci

on, Desarrollo e

Innovaci

on of the Spanish Government (A. Gonz

alez-

opez acknowledges its support under FPI Grant Pro-

gram). The authors would like to thank Hospital do

Barbanza, Ribeira (Spain) for providing images.

REFERENCES

Alonso-Caneiro, D., Read, S. A., and Collins,

M. J. (2013). Automatic segmentation of

choroidal thickness in optical coherence tomog-

raphy. Biomed. Opt. Express, 4(12):2795–2812.

Besag, J. (1974). Spatial interaction and the statistical

analysis of lattice systems. Journal of the Royal

Statistical Society. Series B, 36:192–236.

Bill, A., Sperber, G., and Ujiie, K. (1983). Physiology

of the choroidal vascular bed. Int. Ophthalmol.,

6(2):101–107.

Bol

on-Canedo, V., Snchez-Maroo, N., and Alonso-

Betanzos, A. (2013). A review of feature selec-

tion methods on synthetic data. Knowl. Inf. Syst.,

34(3):483–519.

C¸ esmeli, E. and Wang, D. L. (2001). Texture segmen-

tation using gaussian-markov random ﬁelds and

neural oscillator networks. IEEE Transactions

on Neural Networks, 12(2):394–404.

Danesh, H., Kaﬁeh, R., Rabbani, H., and Hajizadeh,

F. (2014). Segmentation of choroidal boundary

in enhanced depth imaging octs using a multires-

olution texture based modeling in graph cuts.

Computational and Mathematical Methods in

Medicine, 2014(479268).

Dash, M. and Liu, H. (2003). Consistency-based

search in feature selection. Artiﬁcial Intelli-

gence, 151(1-2):155–176.

Daubechies, I. (1992). Ten Lectures on Wavelets.

Society for Industrial and Applied Mathematics,

Philadelphia, PA, USA.

ChoroidCharacterizationinEDIOCTRetinalImagesBasedonTextureAnalysis

275

Dhoot, D., Huo, S., Yuan, A., Xu, D., Srivistava,

S., Ehlers, J., Traboulsi, E., and PK., K. (2013).

Evaluation of choroidal thickness in retinitis pig-

mentosa using enhanced depth imaging optical

coherence tomography. British Journal of Oph-

thalmology, 97(1):66–9.

Gonz

alez, A., Ortega, M., Penedo, M., and Charl

on,

P. (2014). Automatic robust segmentation of reti-

nal layers in oct images with reﬁnement stages.

In International Conference on Image Analysis

and Recognition (ICIAR), volume 8815 of Lec-

ture Notes in Computer Science, pages 337–349.

Springer Berlin Heidelberg.

Hall, M. (1999). Correlation-based Feature Subset

Selection for Machine Learning, PhD disserta-

tion. PhD thesis, University of Waikato.

Haralick, R., Shanmugam, K., and Dinstein, I. (1973).

Textural features for image classiﬁcation. IEEE

Transactions on Systems, Man, and Cybernetics,

3(6):610–621.

Hu, Z., Wu, X., Ouyang, Y., Ouyang, Y., and Sadda,

S. (2013). Semiautomated segmentation of the

choroid in spectral-domain optical coherence to-

mography volume scans. Investigative Ophthal-

mology & Visual Science, 54(3):1722–9.

Imamura, Y., Fujiwara, T., Margolis, R., and Spaide,

R. (2009). Enhanced depth imaging optical co-

herence tomography of the choroid in central

serous chorioretinopathy. Retina, 29(10):1469–

73.

Kaji

c, V., Esmaeelpour, M., Glittenberg, C., Kraus,

M. F., Honegger, J., Othara, R., Binder, S., Fuji-

moto, J. G., and Drexler, W. (2013). Automated

three-dimensional choroidal vessel segmentation

of 3d 1060 nm oct retinal data. Biomed. Opt. Ex-

press, 4(1):134–150.

Lu, H., Boonarpha, N., Kwong, M., and Zheng, Y.

(2013). Automated segmentation of the choroid

in retinal optical coherence tomography images.

In IEEE Eng Med Biol Soc, pages 5869–72.

Mallat, S. G. (1989). A theory for multiresolution

signal decomposition: The wavelet representa-

tion. IEEE Trans. Pattern Anal. Mach. Intell.,

11(7):674–693.

Puzyeyeva, O., Lam, W. C., and Flanagan, J. e. a.

(2011). High-resolution optical coherence to-

mography retinal imaging: A case series illus-

trating potential and limitations. J. Ophthalmol-

ogy, pages 764183–6.

Yin, Z. Q., Vaegan, T. J., Millar, T. J., Beaumont,

P., and Sarks, S. (1997). Widespread choroidal

insufﬁciency in primary open-angle glaucoma.

Journal of Glaucoma, 6(1):23–32.

Zhang, L., Lee, K., Niemeijer, M., Mullins, R.,

Sonka, M., and Abrmoff, M. (2012). Automated

segmentation of the choroid from clinical sd-oct.

Investigative Ophthalmology & Visual Science,

53(12):7510–9.

Zhao, Z. and Liu, H. (2009). Searching for interacting

features in subset selection. Intell. Data Anal.,

13(2):207–228.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

276