Automatic Identiﬁcation of Macular Edema in Optical Coherence

Tomography Images

Gabriela Samagaio

, Aída Estévez

, Joaquim de Moura

1,2

, Jorge Novo

1,2

, Marcos Ortega

1,2

and María Isabel Fernández

3,4,5

Department of Computing, University of A Coruña, A Coruña, Spain

CITIC- Research Center of Information and Communication Technologies, University of A Coruña, A Coruña, Spain

Instituto Oftalmológico Gómez-Ulla, Santiago de Compostela, Spain

Department of Ophthalmology, Complejo Hospitalario Universitario de Santiago, Santiago de Compostela, Spain

University of Santiago de Compostela, Santiago de Compostela, Spain

Keywords:

Computer Aided Diagnosis, Retinal Imaging, Optical Coherence Tomography, Macular Edema.

Abstract:

This paper proposes a novel system for the simultaneous identiﬁcation and characterization of the three types

of Macular Edema (ME) in Optical Coherence Tomography (OCT). These MEs are clinically deﬁned, by the

reference classiﬁcation of the ﬁeld, as: Serous Retinal Detachment (SRD), Diffuse Retinal Thickening (DRT)

and Cystoid Macular Edema (CME). Our system uses multilevel image thresholding approaches to identify

the SRD and CME cases and a learning approach for the DRT identiﬁcation. The system provided promising

results with F-Measures of 83.35% and 81.95% for the DRT and CME detections, respectively. It was also

efﬁcient in detecting all the SRD cases included in the testing image dataset. The system was able to identify

individually the different types of ME on the OCT images but it was also capable to detect simultaneously the

existence of the three ME cases when they appeared merged in the lower retinal layers.

1 INTRODUCTION

According to the World Health Organization, 9.6% of

the European citizens are affected by blindness dise-

ases, situation that is even worse in developing coun-

tries. As reference, in Sub-Saharan Africa, it can re-

ach until 50% of the population. For the last ten years,

cataracts have remained as the main cause of visual

impairment, followed by macular disorders (WHO,

2012). One of the most relevant of them is the Ma-

cular Edema (ME), deﬁned as intraretinal ﬂuid accu-

mulation that affects the central retinal vision, suffe-

ring morphological alterations in the retinal structures

(Trichonas and Kaiser, 2014).

To diagnose these retinal diseases, ophthalmolo-

gists normally support their clinical evaluations in the

analysis of different types of eye fundus images. One

of the most widely used in this ﬁeld is the OCT image

modality. This technique allows a non-invasive and

contactless evaluation of the in vivo histopathology

(Helmy and Allah, 2013). Based on this image mo-

dality, in (Otani et al., 1999), a clinical classiﬁca-

tion was established for the different types of ME

that can be identiﬁed. This classiﬁcation is the re-

ference of the ﬁeld, being clinically used worldwide

by specialists. Intraretinal ﬂuid accumulation was de-

ﬁned in three types based on clinical characteristics

of the images, mainly properties as retinal thickness,

reﬂectivity or area of the abnormalities. These three

types are: Serous Retinal Detachment (SRD), Dif-

fuse Retinal Thickening (DRT) and Cystoid Macu-

lar Edema (CME). Posteriorly, in the work of (Pa-

nozzo et al., 2004), they also characterized each type

by the deﬁnition of 5 parameters: retinal thickness,

diffusion, volume, morphology and epiretinal traction

(Baamonde et al., 2017).

Figure 1 illustrates the presence of the three types

of MEs, where the SRD and the CME cases are hypo-

reﬂective ﬂuid regions with a speciﬁc swollen shape

within the retinal layers, as described by the authors.

Usually, the SRD edemas appear as a dome-shape in

the outer retina while the CME typically appears as

a circular shape in the inner retina, (Joussen et al.,

2010). In the case of the DRT edemas, they are ty-

Samagaio, G., Estévez, A., Moura, J., Novo, J., Ortega, M. and Fernández, M.

Automatic Identiﬁcation of Macular Edema in Optical Coherence Tomography Images.

DOI: 10.5220/0006544105330540

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 4: VISAPP, pages

533-540

ISBN: 978-989-758-290-5

533

pically characterized by a "sponge-like" swelling ap-

pearance that results from the spread of ﬂuid in the

outer retina.

To diagnose these types of diseases, in the recent

years, computational systems have been broadly used

by ophthalmologists as useful tools that allow the di-

agnosis (even in early stages), treatment and moni-

toring of the evolution of the patients. In this ﬁeld,

some works based on the OCT image analysis are

taking initial approaches to help and support the cli-

nical decisions in the analysis of the ME. Therefore,

some efforts have been applied in order to detect the

intraretinal ﬂuid on the OCT images, where particu-

lar characteristics are used as intensity, morphology,

relative position and central and parafoveal retinal

thickness (Willoughby et al., 2017; Montuoro et al.,

2017). At the moment, none of the published works

faced the three types of ME that appear in the macu-

lar region. As reference, in the work of (Sidibé et al.,

2017), the authors proposed a method based on the

Gaussian Mixture Model (GMM) to classify the OCT

scans as normal and abnormal patients. Following a

similar strategy, others (Montuoro et al., 2017) iden-

tiﬁed abnormal OCT images. In this case, they per-

form a simultaneous 3D segmentation of the retinal

layers with the identiﬁcation of two ﬂuid regions as

intraretinal ﬂuid and Sub Retinal Fluid (SRF) using

a graph-theoretic approach. The method of (Alsaih

et al., 2017) used learning strategies in OCT retinal

images in order to identify normal volumes versus vo-

lumes with ME presence. This analysis was based on

the evaluation of the retinal thickening, hard exuda-

tes, intraretinal cystoid space formation, and subre-

tinal ﬂuid. In the case of (González et al., 2013), a

method is proposed to detect the presence of cysts.

A Watershed algorithm is applied within the retinal

tissue in order to ﬁnd all the possible regions in the

image which might conform cystoid structures. Fi-

nally, in order to discard false positives, a learning

strategy is applied to reduce the false positive set. The

authors from (de Moura et al., 2017) proposed a met-

hod to identify the intraretinal cystoid regions, as re-

gions of the OCT images that contain cysts. Hence,

they deﬁned a window size to analyze and extracts a

set of image characteristics to determine the presence

of cysts inside those regions.

In this paper, we propose a novel system to detect

and characterize the intraretinal ﬂuid as SRD, DRT

and CME types, based on the clinical classiﬁcation of

reference in the ﬁeld, the Otani classiﬁcation. As in-

dicated, to date, no other work faced completely the

automatic identiﬁcation of all the types of ME. To ﬁnd

the 3 types of ME, we ﬁrstly delimited the retinal area

in the OCT images, where the intraretinal ﬂuid forms

the swollen regions. Following the Otani ME clini-

cal characterization, we identify the presence of each

type inside this region of interest.

The system will provide help in the standardi-

zation of the identiﬁcation of the different types of

ME, reducing the subjectivity of the ophthalmolo-

gists. Moreover, given the complexity of extraction

of some ME cases, the proposed system will facilitate

the doctor’s work, allowing the early diagnosis and

consequently the procedure with more adjusted treat-

ments, improving the life quality of the patients.

2 METHODOLOGY

The proposed system receives, as input, an OCT re-

tinal image centered in the macula. Firstly, the sy-

stem segments automatically the retinal layers to de-

limit the region of interest (ROI) where the MEs are

present. Inside this region, different strategies were

applied to detect the 3 types of ME: SRD, DRT and

CME, as illustrated in the diagram of Fig. 2.

The system was subdivided in three main steps for

each ME search. Regarding SRD detection, a multi

level adaptive image thresholding was used in order

to ﬁnd candidates with the lower intensity proﬁles.

To discard the false positives detections, different ru-

les based on the clinical knowledge were implemen-

ted. Hence, the candidates should have the speciﬁc

relative position inside the retinal layers as well as a

particular morphological shape (dome-shape). A si-

milar strategy was used to detect the CME cases, as

they also present a deﬁned ovoid shape with a con-

trast with the retinal layers. An image thresholding

method was used to identify CME candidates, follo-

wed by a ﬁltering process of several morphological

conditions as a way of increasing the efﬁciency of the

proposed system. Finally, DRT edemas do not pre-

sent well-deﬁned boundaries and sufﬁcient contrast

with the surrounding tissue. In this way, a learning

strategy was used to distinguish the "sponge-like" re-

gions from the normal regions.

2.1 Retinal Layer Segmentation

Given the noisy characteristics of the input OCT ima-

ges, a median ﬁlter was applied as a pre-processing

step, clearing the image and preserving, simultane-

ously, the properties of the retinal layers in order to

facilitate the posterior identiﬁcations.

In this work, 4 retinal layers were identiﬁed as

they provide the correct delimitation of the regi-

ons where the different types of ME typically ap-

pear. The identiﬁed retinal layers are: Inner Limiting

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

534

Figure 1: Example of OCT retinal image with 3 types of MEs: SRD (+), CME (*) and DRT (- -).

Figure 2: Main methodology steps for the ME identiﬁcation

and characterization in SRD, DRT and CME types.

Figure 3: Example of OCT retinal image with N seeds rand-

omly generated, near the ISOS layer.

Membrane (ILM), Outer Plexiform Layer (OPL), the

junction of Inner and Outer Segments (ISOS) of the

photoreceptors layers and the Retinal Pigment Epit-

helium (RPE).

We used an approach based on the work of (Chiu

et al., 2010) to identify the indicated retinal layers.

In particular, this automatic approach uses graph the-

ory and dynamic programming to represent each OCT

image as a graph of nodes, connecting optimum paths

from both sides of the image. Firstly, the algorithm

calculates dark-to-light gradient images, identifying

adjacent layers and generating weights for the layer

segmentations. The progressive identiﬁcation of the

main layers of the retina was found by the minimum

weighted paths using the (Dijkstra, 1959) algorithm.

This approach detects eight different layers in normal

and healthy OCT retinal images. However, for this

purpose, we adapted the method to identify 3 speciﬁc

layers: ILM, ISOS and RPE.

Regarding the OPL layer, we followed a different

strategy given the deteriorated conditions of the reti-

nal layers that presents the used images of this work.

In order to solve this issue, the previous identi-

ﬁcation of the ISOS layer was used as reference for

the application of region growing (Zhu and Yuille,

1996). Therefore, over this layer N initial points as

seeds were randomly generated, as shown in Fig. 3.

The number of seeds corresponds to a 5% of the in-

put image width. Hence, we use, as baseline, the

ISOS layer to extract the region immediately over it

that corresponds to the OPL layer. Using a signiﬁ-

cant number of seeds along the image we guarantee

the OPL extraction even in signiﬁcantly deteriorated

conditions that present ME cases in advance stages.

2.2 Division in ILM/OPL and

OPL/RPE Regions

As enunciated before, ME consists of the accumula-

tion of ﬂuid within the retinal area. However, each

type of edema typically appears in particular regions

inside the retinal layers that are delimited by the ILM

and RPE retinal layers, as shown in Fig. 4.

According to the Otani classiﬁcation, SRD and

DRT edemas usually appear in the outer retina whe-

reas CME edemas normally start manifesting in the

inner retina, but they can proliferate to the outer retina

in more severe pathological stages, merging with the

DRT cases (Gelfand et al., 2012). Therefore, based

on the previous ROI segmentation that delimits the

retinal tissue, 2 sub-regions were identiﬁed: one cor-

responding to the inner retinal and other for the outer

retina. The inner retina is comprehended between the

ILM and the OPL layers, while the outer retina is deli-

mited between the OPL and the RPE layers, as shown

in Fig. 5. Thus, the detection of each type of intrare-

tinal ﬂuid is simpliﬁed, as the region to ﬁnd each ME

type is reduced.

2.3 SRD Detection

SRD edemas are typically deﬁned as hiporeﬂective

ﬂuid accumulation presenting a dome-shape, with a

characteristic relative position inside the retinal tis-

sue. Therefore, the proposed system was inspired in

these heuristics to automatically identify the SRD pre-

sence.

Firstly, a multi level thresholding, based on the

method of (Otsu, 1979), was applied only in the ou-

Automatic Identiﬁcation of Macular Edema in Optical Coherence Tomography Images

535

Figure 4: Example of OCT retinal image. (a) OCT image

with the presence of CME edemas. (b) Binary image mask

with the delimitation of the ROI between the ILM and the

RPE retinal layers.

Figure 5: Inner and Outer regions of interest. (a) Entire

ROI. (b) Inner retina between the ILM and the OPL retinal

layers. (c) Outer retina between the OPL and RPE retinal

layers.

ter retina. With this thresholding, we identify the re-

gions with the lower intensity proﬁles as candidates

for being SRD edema. For that, the optimal thres-

hold was determined with the value that maximizes

the separability between the intraretinal ﬂuid regions

and the retinal layers (Noma et al., 2011; Gupta et al.,

2013).

Next, the objective is the removal of those candi-

dates that do not fulﬁll the medical restrictions that

characterize SRD edemas. Following this medical re-

strictions, several conditions were implemented ensu-

ring that the system detects efﬁciently the SRD ede-

mas. These conditions are:

• Relative position. This type of edemas should stay

near the photoreceptors layer, (Carmona and Her-

nández, 2015);

• Area size, bigger than 200µm

. We selected, as

reference, the area of the microcystic macular

edema (Gelfand et al., 2012; Wolff et al., 2014);

• Fusiform morphology. The width that should be

within an empirical range, [200 - 980]µm equiva-

lent to [51 - 250] pixels, the typical lengths that

they normally present;

• Constriction of the photoreceptors region. The

SRD leads to a decrease of the region thickness

delimited between the ISOS and the RPE layers in

the parafoveal zone (Ooto et al., 2010). Therefore,

Figure 6: Example of OCT with the ﬁnal ROI where was

removed the region of the SRD edema (+) and the photore-

ceptor layer (- -).

near the fovea, we compared the mean thickness

of the empirical window size and the global mean

thickness of the ISOS/RPE region. The result

should be less than 1 to consider the presence of

candidates for being SRD edemas;

• The intensity of the area above SRD edema. This

region should have a higher intensity proﬁle when

compared with the ﬂuid region, (Ooto et al.,

2010). This occurs because of part of the photo-

receptors layer usually stays above the SRD ede-

mas, as a brighter region. Hence, the mean inten-

sity of the candidate is compared with the mean

intensity of the photoreceptors layer, using the

same empirical window size.

As SRD edema, if present, only appears one per

image, if two or more candidates fulﬁll all these con-

ditions, the system will preserve the candidate that is

more centered, near the fovea.

Finally, using the SRD detection as seed, a region

growing is employed to obtain the SRD segmentation

as more precise as possible. This precise extraction is

useful in posterior stages of the method. An example

of the SRD segmentation is presented in Fig. 6.

2.4 DRT Detection

DRT or "sponge-like" edema is deﬁned by the speci-

alists as a retinal swelling of the macula with reduced

intraretinal reﬂectivity. Also, this edema is typically

located in the inner retina, being usually above the

photoreceptor layer. Therefore, to ﬁnd this type of

intraretinal ﬂuid accumulation, the proposed system

searches for the DRT appearance in the OPL/RPE re-

gion, but removing the photoreceptors region and the

SRD edema, when it is detected. The precise ex-

traction of the SRD edema in the previous stage faci-

litates the SRD removal to calculate the new region of

interest. Hence, the correction of the OPL/RPE region

can be achieved accurately, resulting in a new region

equivalent to the OPL/ISOS, decreasing the detection

of false candidates with a more precise and restricted

region to detect the DRT presence.

As this type of edema does not have a well-deﬁned

morphological shape and contour, simple methods ba-

sed on image processing techniques are not sufﬁcient

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

536

Figure 7: Example of OCT image with the new baseline on

the OPL/ISOS region. The Naive Bayes classiﬁer is trained

by the feature extraction from each "esponge-like" columns.

to produce acceptable results. Therefore, in this step,

a learning strategy was implemented using the Naive

Bayes classiﬁer, extracting features per column from

the OPL/RPE region and identifying the DRT pre-

sence, as illustrated in the diagram of Fig. 7. We

used Naive Bayes as a frequently used classiﬁer in

medical imaging approaches. The extracted featu-

res were based mainly on intensity and texture featu-

res. Moreover, another relevant property is the retinal

thickness, as MEs normally produce the ﬂuid accu-

mulation inside the retinal layers and, therefore, the

increment of their thickness (Goebel and Kretzchmar-

Gross, 2002). We used this property also as an indi-

cator of the disease presence.

Table 1 details the 18 features that were imple-

mented to identify the DRT presence. A Sequential

Forward Selection (SFS) method was applied to the

feature set to reduce the array dimensionality and pre-

serve those with the highest discriminative power.

Table 1: List of the 18 features to identify the DRT pre-

sence.

of Features Feature

- Intensity Image Analysis:

1 - 6 Maximum, Minimum,

Mean, Median,

Standard-deviation and Variance

- Histogram Image Analysis:

7 - 10 Obliquity, Kurtosis,

Energy and Entropy

- Mask Height Analysis:

Height of OPL/RPE region,

11 -14 Height of mask ILM/RPE

The ratio between the heights of

OPL/RPE and the ILM/RPE re-

gions

- Texture Analysis:

15 -18 GLCM method: Contrast,

Correlation, Energy and

Homogeneity

2.5 CME Detection

Cysts typically present a low intensity proﬁle with a

signiﬁcant contrast with the ILM/OPL region, where

they frequently appear. A multilevel thresholding, ba-

sed on the Otsu’s algorithm, was also implemented in

(a)

(b)

(c)

Figure 8: Example of OCT image with the presence of

CMEs. (a) Input OCT retinal image. (b) CMEs represented

as (*) in the ILM/OPL region. (c) CMEs (*) in OPL/RPE

region with lower contrast compared to the surrounded reti-

nal layer.

this stage, due to the contrast between the ILM/OPL

region and the CMEs, identifying all the candidates

for being cysts. Due to the relative high number of

detected false positives, a post processing stage was

implemented to preserve the real existing ones. As we

can retrieve partial segmentations (even several candi-

dates for the same cyst), we completed the segmenta-

tions using Watershed. This way, we obtain more de-

ﬁned reliable regions for the candidates that facilitate

the posterior analysis and the false positive reduction.

Hence, using as reference the clinical properties that

the specialists follow to identify the cysts, we imple-

mented a list of conditions to preserve the real CMEs

from the list of candidates (Wolff et al., 2014; Joussen

et al., 2010; Helmy and Allah, 2013). These conditi-

ons are listed below:

• Area size, bigger than 1µm

. To discard noisy and

tissue artifacts;

• Eccentricity, should be smaller than 0.98 to re-

spect the ovoid shape. This parameter is the ra-

tio of the distance between the central point of the

ellipse and its major axis length;

• Retinal thickness, the ROI (ILM/RPE) thickness

should be bigger than 250µm equivalent to 64

pixels. This value corresponds to the nor-

mal size of the parafoveal region, (Goebel and

Kretzchmar-Gross, 2002);

• The width of the candidate area should be inside

a range of sizes [40 - 530]µm equivalent to [10

- 135] pixels. This way, the range ensures that

the area of the candidates should be able to re-

move small artifacts and regions with a length that

is approximately twice the normal of the retinal

thickness.

Automatic Identiﬁcation of Macular Edema in Optical Coherence Tomography Images

537

Basically, the restrictions are based in terms of area,

eccentricity, retinal layer thickness as well as the

thickness between the speciﬁc retinal layers. As in-

dicated before, the presence of CMEs leads to an in-

crease of the retinal layers where the ﬂuid is accumu-

lated, and consequently, produce a global increment

of the retinal tissue (ILM/RPE region).

As Fig. 8 shows, it is possible to conclude that

the large majority of CMEs are within the ILM/OPL

region. However, in more advanced pathological sta-

ges, they can also proliferate to the OPL/RPE region.

In the outer region the contrast is lower, which makes

the identiﬁcation of the existing CMEs a more com-

plex issue. In this case, we followed the same strategy

as before, but adapting the parameters to the new con-

ditions.

3 RESULTS AND DISCUSSION

As presented, the proposed system includes three dif-

ferent strategies to identify each ME type. All the

strategies were tested using an image dataset that is

composed by 50 OCT retinal images, centered in the

macula, with a resolution of 2032 × 596 pixels. This

dataset was acquired with a Spectralis

 OCT confo-

cal scanning laser ophthalmoscope from Heidelberg

Engineering.

To ensure the efﬁciency of the system all the ima-

ges were labeled by an expert clinician, identifying

the location of the 3 types of MEs inside each scan.

Based on this ground-truth, we constructed the trai-

ning and testing sets for the DRT identiﬁcation ap-

proach as well as validated the performance of SRD

and CME identiﬁcation approaches.

Precision, Recall and F-Measure were the used

metrics for the validation of the proposed system, as

indicated in Eqs.1, 2 and 3, respectively. The F-

Measure is deﬁned as a combination of both precision

and recall metrics in a global measurement.

Precision =

T P

T P + FP

(1)

Recall =

T P

T P + FN

(2)

F − Measure = 2 ∗

Precision ∗ Recall

Precision + Recall

(3)

where TP are the True Positives, FP the False Positi-

ves and FN the False Negatives.

Regarding the deﬁned parameters, we used values

that were empirically calculated, as it was previously

mentioned in the methodology.

The efﬁciency of the proposed method for ME in

OCT retinal images was evaluated according to quan-

titative metrics. As gold standard, for the SRD and

CME cases, we measured if the central point of each

detected edema was successfully identiﬁed, based on

the specialist segmentation. Regarding the DRT case,

we analyzed if each column was correctly identiﬁed

when compared with the specialist opinion.

Regarding the case of SRD edemas, they are not

as common as the others, affecting only a reduced

groups of patients, as presented in (Otani et al., 1999).

Moreover, when it is present, only one SRD patho-

logical structure can be identiﬁed in each OCT scan.

For that reason, the employed image dataset only con-

tains 4 SRD edemas. The 4 cases were correctly iden-

tiﬁed by the proposed system.

For the DRT detection, the Naive Bayes classiﬁer

was trained using the proposed dataset with a 10 fold

cross validation. Per each OCT image in the data-

set, 80 samples were randomly selected, representing

equally both DRT and non-DRT cases, resulting in a

total of 4000 extracted samples.

Regarding the selected features using the SFS

method, 4 of them were taken from the initial set:

mean intensity, kurtosis, energy and energy from the

GLCM matrix, as they include a high discriminant

power to differentiate common retinal tissue patterns

with respect to the DRT presence.

Using the selected features, the Naive Bayes clas-

siﬁer was trained and tested satisfactorily providing

the presented metrics resulting in Table 2.

Table 2: Performance of the DRT detection approach.

Precision Recall F-Measure

84.04% 80.79% 83.35%

With this strategy, we were able to detect the

83.35% of the total regions of DRTs. The introduced

mistakes are mainly derived from shadows that are

produced by the presence of vessels but also from dif-

ferent pathological structures (hard-exudates). These

shadows change the typical characteristics of the DRT

edemas. Therefore, these artifacts lead to a miss-

classiﬁcation of the "sponge-region", decreasing the

metric.

Regarding the CME case, it can be found in both

regions (inner and outer retina), simultaneously. The-

refore, the efﬁciency of the system with this ME type

was tested in 2 phases, as listed in Table 3. In the ﬁrst

phase, the system was tested in the inner retina where

the results were better. The F-Measure reaches a va-

lue of 87.48% in this case. The errors are mainly deri-

ved from missing CMEs with speciﬁc morphologies,

as fusiform shapes and the presence of Microcystic

Macular Edemas (MME). In the ﬁrst case, cysts exhi-

bit an unusual elongated shape, in the horizontal axis.

Whereas, as a second case, the MMEs have a small

area with a not well deﬁned boundaries.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

538

Figure 9: Example of OCT retinal image with undetected

MMEs (x) and correct detections of CMEs (*).

Figure 10: Example of the OCT retinal non detected CMEs

(- -), due to the fusiform shape.

In the second phase, the method was tested in the

outer retina where lower results were obtained given

the higher complexity of the identiﬁcation in this re-

gion. This decrease of the metrics is, as commented

before, due to the fact that the almost absent contrast

between the CME and the retina tissue. In a more

complex scenario, the method is also able to detect

efﬁciently the CMEs in the OPL/RPE region.

Figure 11 presents an illustrative result where the

proposed system detects and characterizes the three

types of ME even when they are simultaneously pre-

sent on the same OCT image.

Table 3: Results of the quantitative metrics Precision, Re-

call and F-Measure for the CME detection.

Region Precision Recall F-Measure

ILM/OPL 96.84% 78.34% 87.48%

OPL/RPE 88.26% 78.19% 74.40%

Both 89.98% 80.34% 81.95%

4 CONCLUSION

In this paper, we propose a novel automatic system

that detects and characterizes the intraretinal ﬂuid in

the different types of ME using OCT retinal ima-

ges. Using clinical criterions inspired in the Otani

classiﬁcation, the system automatically identiﬁes and

characterizes the ME in three types: SRD, DRT and

CME. To do that, we analyzed intensity, texture, mor-

phological and position properties of them retinal tis-

sue to detect all the cases of each ME type. Regarding

SRD and CME edemas, different approaches were

implemented based on adaptive thresholding in speci-

ﬁc regions of the retinal layers, as they usually present

identiﬁable boundaries and a signiﬁcant contrast with

the retinal tissue. Then, to remove FPs, clinical con-

ditions were applied based on the typical morphology

(a)

(b)

Figure 11: Example of an OCT retinal image with the de-

tection of the three types of macular edema: SRD illustrated

as (+), CME as (*) and DRT the selected columns.

of each type of ME. In the case of the DRT detection,

a Naive Bayes classiﬁer was trained to identify the re-

gions where it is present using a list of 18 features as

intensity, texture and domain knowledge properties.

Then SFS is applied in order to select the ones with

higher power of discrimination.

Experimental results show that the proposed sy-

stem achieved promising results for the ME identi-

ﬁcation and characterization, even when they appear

combined on the same retinal region. In particular,

our system achieves a F-Measure of 83.35% for the

DRT and 81.95% in the CME detection. While, in

the detection of the SRD, the system detects all ede-

mas that were present in the used dataset. Therefore,

this system can be an important tool in clinical ﬁelds

helping in the detection and characterization of retinal

diseases.

As future works, the proposed system could have

a larger dataset in order to reinforce the conclusions

that were achieved in this work. Moreover, we aim

the identiﬁcation of more complex cases of the CME

edemas. We also want to detect microcystic macular

edema. For doing that, speciﬁc approaches can be de-

signed with that purpose. Using the identiﬁed MEs,

different statistic metrics can be derived as a way to

provide valuable information to specialists and faci-

litate their diagnostic process. Therefore, it will be

possible to motorize the disease evolution and incre-

ase the life quality of the patients.

ACKNOWLEDGEMENTS

This work is supported by the Instituto de Salud Car-

los III, Government of Spain and FEDER funds of

Automatic Identiﬁcation of Macular Edema in Optical Coherence Tomography Images

539

the European Union through the PI14/02161 and the

DTS15/00153 research projects and by the Ministerio

de Economía y Competitividad, Government of Spain

through the DPI2015-69948-R research project. Also,

this work has received ﬁnancial support from the Eu-

ropean Union (European Regional Development Fund

- ERDF) and the Xunta de Galicia, Centro singular

de investigación de Galicia accreditation 2016-2019,

Ref. ED431G/01; and Grupos de Referencia Compe-

titiva, Ref. ED431C 2016-047.

REFERENCES

Alsaih, K., Lemaitre, G., Rastgoo, M., Massich, J., Sidibé,

D., and Meriaudeau, F. (2017). Machine Learning

Techniques for Diabetic Macular Edema (DME) Clas-

siﬁcation on SD-OCT Images. Biomed. Eng. Online,

16(1):68.

Baamonde, S., de Moura, J., Novo, J., and Ortega, M.

(2017). Automatic Detection of Epiretinal Membrane

in OCT Images by Means of Local Luminosity Pat-

terns. In International Work-Conference on Artiﬁcial

Neural Networks, pages 222–235. Springer.

Carmona, L. and Hernández, F. (2015). Revisión Biblio-

gráﬁca: Edema Macular Diabético, Repercusiones y

Tratamiento. Revista Médica del Instituto Mexicano

del Seguro Social, 53(5):600–607.

Chiu, S., Li, X., Nicholas, P., Toth, C., Izatt, J., and Farsiu,

S. (2010). Automatic Segmentation of Seven Reti-

nal Layers in SDOCT Images Congruent with Expert

Manual Segmentation. Opt. Express, 18(18):19413–

19428.

de Moura, J., Novo, J., Rouco, J., Penedo, M., and Ortega,

M. (2017). Automatic Identiﬁcation of Intraretinal

Cystoid Regions in Optical Coherence Tomography.

In Conference on Artiﬁcial Intelligence in Medicine in

Europe, pages 305–315. Springer.

Dijkstra, E. (1959). A Note on Two Problems in Connexion

with Graphs. Numerische Mathematik, 1(1):269.

Gelfand, J., Nolan, R., Schwartz, D., Graves, J., and Green,

A. (2012). Microcystic Macular Oedema in Multiple

Sclerosis is Associated with Disease Severity. Brain,

135(6):1786–1793.

Goebel, W. and Kretzchmar-Gross, T. (2002). Retinal

Thickness in Diabetic Retinopathy: A Study Using

Optical Coherence Tomography (OCT). Retina,

22(6):759–767.

González, A., Remeseiro, B., Ortega, M., Penedo, M., and

Charlón, P. (2013). Automatic Cyst Detection in OCT

Retinal Images Combining Region Flooding and Tex-

ture Analysis. In CBMS, 2013 IEEE 26th Internatio-

nal Symposium on, pages 397–400. IEEE.

Gupta, A., Raman, R., Mohana, K., Kulothungan, V., and

Sharma, T. (2013). Communications Between Intra-

retinal and Subretinal Space on Optical Coherence

Tomography of Neurosensory Retinal Detachment in

Diabetic Macular Edema. Oman J.of Ophthalmol.,

6(3):183.

Helmy, Y. and Allah, H. (2013). Optical Coherence To-

mography Classiﬁcation of Diabetic Cystoid Macular

Edema. Clin. Ophthalmol., 7:1731.

Joussen, A., Gardner, T., Kirchhof, B., and Ryan, S. (2010).

Retinal Vascular Disease. Springer Science & Busi-

ness Media.

Montuoro, A., Waldstein, S., Gerendas, B., Schmidt-

Erfurth, U., and Bogunovi

c, H. (2017). Joint Reti-

nal Layer and Fluid Segmentation in OCT Scans of

Eyes with Severe Macular Edema Using Unsupervi-

sed Representation and Auto-Context. Biomed. Opt.

Express, 8(3):1874–1888.

Noma, H., Funatsu, H., Mimura, T., and Shimada, K.

(2011). Visual Function and Serous Retinal Deta-

chment in Patients with Branch Retinal Vein Occlu-

sion and Macular Edema: A Case Series. BMC Opht-

halmology, 11(1):29.

Ooto, S., Tsujikawa, A., Mori, S., Tamura, H., Yamashiro,

K., and Yoshimura, N. (2010). Thickness of Photore-

ceptor Layers in Polypoidal Choroidal Vasculopathy

and Central Serous Chorioretinopathy. Graefe’s Arch.

Clin. and Exp. Ophthalmol., 248(8):1077–1086.

Otani, T., Kishi, S., and Maruyama, Y. (1999). Patterns

of Diabetic Macular Edema with Optical Coherence

Tomography. Am. J. of Ophthalmol., 127(6):688–693.

Otsu, N. (1979). A Threshold Selection Method from Gray-

Level Histograms. EEE Trans. Syst. Man. Cybern B.

Cybern., 9(1):62–66.

Panozzo, G., Parolini, B., Gusson, E., Mercanti, A., Pi-

nackatt, S., Bertoldo, G., and Pignatto, S. (2004). Dia-

betic Macular Edema: An OCT-Based Classiﬁcation.

In Seminars in Ophthalmology, volume 19, pages 13–

20. Taylor & Francis.

Sidibé, D., Sankar, S., Lemaître, G., Rastgoo, M., Mas-

sich, J., Cheung, C., Tan, G., Milea, et al. (2017).

An Anomaly Detection Approach for the Identiﬁca-

tion of DME Patients Using Spectral Domain Optical

Coherence Tomography Images. Comput. Methods

and Programs in Biomed., 139:109–117.

Trichonas, G. and Kaiser, P. (2014). Optical Coherence To-

mography Imaging of Macular Oedema. Br. J. of Op-

hthalmol., 98(Suppl 2):ii24–ii29.

WHO (2012). Global Data on Visual Impairments 2010.

Geneva: World Health Organ.

Willoughby, A., Chiu, S., Silverman, R., Farsiu, S., Bai-

ley, C., Wiley, H., Ferris, F., and Jaffe, G. (2017).

Platform-Independent Cirrus and Spectralis Thickness

Measurements in Eyes with Diabetic Macular Edema

Using Fully Automated Software. Transl. Vis. Sci. &

Technol., 6(1):9–9.

Wolff, B., Azar, G., Vasseur, V., Sahel, J.-A., Vignal, C.,

and Mauget-Faÿsse, M. (2014). Microcystic Changes

in the Retinal Internal Nuclear Layer Associated with

Optic Atrophy: a Prospective Study. J. Ophthalmol.,

2014.

Zhu, S. and Yuille, A. (1996). Region Competition: Uni-

fying Snakes, Region Growing, and bayes/MDL for

Multiband Image Segmentation. IEEE Trans.s on Pat-

tern Anal. and Mach. Intell., 18(9):884–900.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

540