2020 International Conference on Machine Vision and Image Processing (MVIP)

Faculty of Engineering, College of Farabi, University of Tehran, Iran, 19 & 20 Feb. 2020

Fully Convolutional Networks for Fluid

Segmentation in Retina Images
Behnam Azimi Abdolreza Rashno Sadegh Fadaei
Department of Computer Engineering Department of Computer Engineering Department of Electrical Engineering
Faculty of Engineering Faculty of Engineering Faculty of Engineering
Lorestan University Lorestan University Yasouj University
Khorramabad, Iran Khorramabad, Iran, 6815144316 Yasouj, Iran, 7591874934
behnam3azimi@gmail.com rashno.a@lu.ac.ir s.fadaei@yu.ac.ir

Abstract—Retinal diseases can be manifested in optical co-
herence tomography (OCT) images since many signs of retina
abnormalities are visible in OCT. Fluid regions can reveal the
signs of age-related macular degeneration (AMD) and diabetic
macular edema (DME) diseases and automatic segmentation
of these regions can help ophthalmologists for diagnosis and
treatment. This work presents a fully-automated method based
on graph shortest path layer segmentation and fully convolu-
tional networks (FCNs) for fluid segmentation. The proposed
method has been evaluated on a dataset containing 600 OCT
scans of 24 subjects. Results showed that the proposed FCN
model outperforms 3 existing fluid segmentation methods by the
improvement of 4.44% and 6.28% with respect to dice cofficients
and sensitivity, respectively.
Index Terms—Fully Convolutional Networks, Fluid Segmenta-
tion, graph shortest path, Retina, Optical Coherence Tomogra-
phy.
I. I NTRODUCTION
Machine learning methods have been applied in many
applications such as image and audio processing for verifi-
cation, identification, recognition and classification [1]–[8]. In
recent years, a concept referred as neutrosophic sets was also
used in mentioned applications [9]–[15]. Ophthalmologists use
machine learning methods to evaluate retina condition in the
term of being in normal or abnormal conditions. For this task,
first images are taken from retina by imaging methodologies
such as fundus and optical coherence tomography (OCT).
Then, structure of retinal layers, blood vessels, tissue regions
area, optic disk and drusen regions can all be computed by
algorithms based on image processing and machine vision.
Since manually extraction of such parameters is a very time-
consuming task, automated algorithms are used for these tasks.
One of the most significant parameters in retina is fluid
regions which are created due to the leakage of blood vessels.
Fluid regions can be detected in OCT scans and are used
as good signs for age-related macular degeneration (AMD)
and diabetic macular edema (DME). Therefore, the area and
volume of fluid regions can be used by ophthalmologists for
treatment and follow-up of retinal diseases.
Many fluid segmentation methods have been proposed in
the literature. Prior to this work, we have presented fully-
automated methods for fluid segmentation and detection as
well as retinal layer segmentation based on graph cut and
graph shorters path methods in Neutrosophic (NS) domain
for both AMD and DME subjects in [16]–[22]. A compre-
hensive fluid detection and segmentation challenge was done
by Bogunovic for three types of fluid [23]. Fluid regions
in OCT scans were segmented and detected by deep neural
networks in [24]. Layers and fluid regions of retinal were
automatically segmented by fully convolutional networks in
a method referred as Relaynet [25]. Convolutional neural
networks were applied for robust retina thickness segmentation
of OCT scans in [26]. Fully-automated quantification and
detection of macular fluid in oct was proposed with deep
learning methods in [27]. Multi-vendor OCT scans were ana-
lyzed by deep learning approaches in the term of intraretinal
cystoid fluid detection and quantification in [28]. Generally,
fluid segmentation can be done by either image segmentation
or pixel classification. In image segmentation methodologies,
the whole image is segmented to fluid and tissue segments
(regions) while in pixel classification, each pixel is classified to
fluid or tissue. For pixel classification, texture, color and shape
features are extrected from pixels and used in conventional
pixel classification methods [29], [30]. Here, fluid regions are
segmented by pixel classification.
The main contribution of this research is to propose a fully-
automated method consisting layer segmentation and fully
convolutional network (FCN) for fluid segmentation in sub-
jects with retina abnormalities. First, inner limiting membrane
(ILM) and retinal pigmentation epithelium (RPE) layers as
first and last layers of retina, respectively, are segmented by
graph shortest path algorithms in NS domain. Then, retinal
region between these layers as region of interest (ROI) is
extracted. OCT scan and its corresponding ROI are presented
as two input channels for FCN. The proposed model for
FCN has two encoder and decoder steps with the depth of
4. In each depth of encoder, two convolution layers and
one max-pooling layer; and in each depth of decoder, two
convolution layers, one transposed convolution layer and one
concatenate layer are used. The final output of FCN is set
to 2 channels as same as input dimension, one for fluid and
another for tissue regions. The rest of this paper is organized

978-1-7281-6832-6/20/$31.00 2020
c IEEE

Authorized licensed use limited to: University of Exeter. Downloaded on June 19,2020 at 23:56:57 UTC from IEEE Xplore. Restrictions apply.
as follows: Section II presents the proposed method including
ROI Segmentation and FCN model. Dataset is described in
Section III. Experimental setup and results are discussed in
Sections IV and V, respectively. Limitations of the proposed
FCN model are explanied by examples in Section VI. Finally,
this work is concluded in Section VII.
II. P ROPOSED METHOD
A. Region of Interest Segmentation
If the first and last layers of OCT scans; named as ILM and
RPE; are segmented, it leads to ROI segmentation and back-
ground elimination. For this task, images are first transferred
to NS domain by Eqs. (1)-(5):
g(i, j) − gmin
T (i, j) =
gmax − gmin
F (i, j) = 1 − T (i, j);
δ(i, j) − δmin
I(i, j) = ; (3)
δmax − δmin
w w
1 
2 
2 convolution layers which are 64, 256 and 512 for depths 2,
g(i, j) = 2 g(i + m, j + n)
w w w
m=− 2 n=− 2
δ(i, j) = |g(i, j) − g(i, j)|
where g is an input image, gmin and gmax are the minimum
and maximum gray level of input image, respectively. T , I and
F represent true, indeterminacy and false sets, respectively.
Then, a graph is created from image by mapping each pixel
in image to one vertex in graph. Edges between vertexes are
computed by gray-levels of pixels connected to those vertexes.
For ILM segmentation, weights are computed by Eqs. (6)-(7)
which is applied to T set in NS domain. .
⎤ ⎡
−2
V erGrad = T ∗ H, H = ⎣ 0 ⎦
2 for segmentation, a convolution layer with 2 filters are used
W ((a1 , b1 ), (a2 , b2 )) = 4 × M axG − V erGrad(a1 , b1 )
− V erGrad(a2 , b2 ) + 2 × mean(R)
(7)
where R is a set of pixels above (a1 , b1 ) and M axG is the
maximum gradient in image. For RPE segmentation, filter H
is inversed and R corresponds to a set of pixels under (a1, b1)
[17].
Two pixels(vertexes) in OCT images including up-left and
down-right pixels are selected as start and end points, respec-
tively. In the next stage, Dijkstra algorithm is applied to find
the shortest path between start and end points which leads to
segment ILM and RPE [17].
ROI segmentation has some advantages as follows: 1) All
fluid regions are located between these layers and no fluid
Authorized licensed use limited to: University of Exeter. Downloaded on June 19,2020 at 23:56:57 UTC from IEEE Xplore. Restrictions apply.
region is lost in segmentation process. 2) Fluid regions are very
similar to background in both brightness and texture lead to
mislead the segmentation method. 3) ROI region is processed
instead of whole image which results in speeding up. The
proposed scheme in this research uses the first and scened
mentioned advantages of ROI segmentation.
B. Fully Convolutional Network Model
Two channels of OCT images including main OCT scan and
its ROI are presented for FCN. Therefore, input dimension is
256x512. The proposed FCN structure for fluid segmentation
is shown in Fig. 1 which consists two encoder and decoder
steps. The depth of both decoder and encoder is 4. There
is no need for further depths due to the input dimension.
In each depth of encoder, two convolution layers with 32
(1)
filters are applied to the whole depth of input channels.
To preserve the input dimension, padding is applied to all
(2) convolution layers which leads to the channels with dimension
of 256x512x32. Then max-pooling layer for downsampling
with window size [2,2] and stride 2 is applied to decrease
the input dimension by half (128x256x32). Other depth of
encoder are repeated by the same way except the number of
(4) 3 and 4, respectively. After encoder steps, two convolution
layers with 512 filters are applied which leads to the dimension
of (16x32x512). In each level of decoder phase, transposed
(5) convolution layer with learning parameters is first used for
upsampling which creates an output as twice as input size.
This layer inverses max-pooling operation. Then, the output
of convolution layer in each level of encoder is sent for
corresponding layer in decoder via skip connection and is
concatenated with the output of transposed convolution in
concatenate layer. For the first level, channels of dimension
32x64x512 are appeared. Finally, two convolution layer with
256 filters and padding are applied which leads to 32x64x256.
In other levels 2, 3 and 4 of decoder; 128, 64 and 32 filters
are used, respectively. By applying transposed convolution in
each level for doubling input size, the final dimension of
(6) decoder is 256x512x32. Finally, since there are two classes
to create final dimension 256x512x2 for FCN. One channel is
considered for tissue revelation and another for fluid pixels.
III. OCT DATASET
To evaluate the effectiveness of the proposed method, a
publicly available OCT dataset containing 600 OCT scans of
24 subjects with the dimension of 496x1024 collected from
University of Minnesota is used. The width and length reso-
lution of OCT scans are 3.87μm/pixel and 5.88μm/pixel, re-
spectively. Ground truth images for fluid regions are available
from two ophthalmologist experts. Among all 600 images, 48
scans are selected for training set and 83 scans for validation.
Therefore, totally 141 scans are used in training phase and
459 scans for testing. The reason for few number of training
scans is that, in training set, all types of fluids including sub-
 Fig. 1: Proposed structure for fluid segmentation by FCN model.
retinal, intra-retinal and sub-RPE are considered and similar
points are ignored.
IV. E XPERIMENTAL SETUP
Because of memory and computation resource limitations,
OCT scans are first resized by half to 256x512. After, fluid
segmentation, they are resized to their original size and
then segmented fluid regions are compared with ground truth
images. Kernel size [3, 3] is used for all convolution layers
followed by LeakyReLu activation function with alpha = 0.2.
Finally, loss function cross entropy and ADAM optimizer are
used. By these configurations, FCN model is trained with
labels of expert 1 and tested against expert 1 and 2 labels
for fluid regions.
V. R ESULTS
Segmented fluid regions by the proposed model are com-
pared with neutrosophic graph cut (NSGC) [17], graph cut
(GC) [32] and kernel graph cut (KGC) [33] methods with
respect to dice coefficient and sensitivity measures. These
methods have been applied to the mentioned OCT dataset and
this is the main reason that why we have compared the results
of our model with them. These measures are computed for
comparison between automated methods in one side and expert
1 and expert 2 in another side. Fig. 2 depicts segmented scans
by the proposed FCN model for subjects with intra-retinal
fluid regions. Note that this type of fluid is revealed between
retinal layers. Segmented scans for sub-retinal and sub-RPE
fluid regions are shown in Figs. 3 and 4, respectively.
Numeric results for segmented fluid regions by Graph Cut
(GC), Kernel Graph Cut (KGC) and Neutrosophic Graph
Cut (NSGC) for dice coefficient and precision measures are
Authorized licensed use limited to: University of Exeter. Downloaded on June 19,2020 at 23:56:57 UTC from IEEE Xplore. Restrictions apply.
reported in Tables I and II, respectively. In each column, a
measure is computed with respect to expert 1 (E.1) and expert
2 (E.2). For dice coefficient measure, the proposed FCN model
outperforms NSGC, KGC and GC in subjects 1, 4, 6, 8-13, 15,
17 and 19-23 in both E.1 and E.2 columns. NSGC achieved
the best dice coefficient in E.1 column of subjects 2, 3, 7 and
14; E.2 column of subject 16; and both E.1 and E.2 columns
of subjects 18 and 24. The proposed method has the best
average dice coefficient of 86% in comparison with NSGC,
KGC and GC with 81.56%, 65.25% and 69.13% respectively.
The proposed method also achieves the best sensitivity in both
E.1 and E.2 columns of subjects 1, 2-6, 8-17, 19 and 21-24.
NSGC has better performance in E.1 column of subjects 2, 7,
18 and 20. Averagely, the proposed model with the sensitivity
of 87.38% outperforms NSGC, KGC and GC with 81.10%,
75.69% and 70.11%, respectively.
VI. D ISCUSSION
Despite the good segmentation results of the proposed FCN
model, there are some issues that affect the FCN model and
leads to bad segmentation results. The first one is errors raised
from inaccurate ROI segmentation. Basically, there is not
any fluid region above ILM and RPE layers. If these layers
are segmented wrongly, fluid regions can be located out of
ROI. All fluid regions out of ROI (above segmented ILM
and beneath segmented RPE) will be missed by FCN since
these regions are trained as tissue. One sample of such cases
is demonstrated in Fig. 5. In this case, fluid region beneath
segmented RPE is missed because of inaccurate segmentation
of RPE.
Also, in some cases with sub-retinal fluid regions, non-fluid
dark regions are appeared above fluid regions and mislead
TABLE I: Dice coefficient of the proposed methods and other
methods.
Dice Coefficients
GC KGC NSGraph Prop. Method
sub E.1 E.2 E.1 E.2 E.1 E.2 E.1 E.2
1 78.68 75.34 79.07 74.76 85.68 81.81 94.41 94.84
2 77.79 77.83 86.11 82.03 89.49 85.36 89.19 91.78
3 56.54 53.05 56.56 51.91 85.81 79.4 84.87 83.75
4 71.94 72.26 69.51 69.49 82.81 82.34 84.45 83.6
5 45.80 43.7 25.2 24.64 68.25 66.02 75.8 72.82
6 73.32 68.61 6134 57.1 85.91 80.53 93.92 94.08
7 56.71 52.61 46.32 42.0 67.31 63.13 54.84 58.04
8 86.89 86.25 79.06 79.07 86.85 86.29 95.79 95.54
9 76.65. 72.61 74.95 70.92 83.0 78.81 95.33 95.24
10 76.54 76.39 76.57 76.44 84.63 80.51 90.01 89.81
11 72.22 67.74 55.28 51.02 86.51 82.26 92.42 87.17
12 71.34 68.03 69.53 66.08 82.42 79.43 90.41 85.87
13 62.61 62.57 64.67 64.56 89.9 89.85 92.28 92.19
14 63.48 59.26 57.34 53.47 81.07 76.98 78.23 78.19
15 61.94 53.73 58.19 50.26 74.16 66.07 78.74 76.48
16 71.17 71.07 69.21 69.2 87.62 87.57 91.76 85.89
17 73.03 72.17 88.31 83.58 91.52 86.48 92.16 91.45
18 66.27 58.43 66.93 59.49 84.7 77.06 76.9 74.66
19 73.93 73.64 43.31 43.26 86.38 82.09 90.54 85.94
20 65.76 57.41 69.75 63.05 85.4 75.13 85.64 79.79
21 81.11 70.62 76.04 66.16 89.0 78.38 93.27 83.03
22 71.33 64.40 61.5 55.0 77.16 70.19 82.13 82.08
23 89.83 79.47 90.85 82.49 91.68 82.66 96.45 87.67
24 81.37 77.12 87.27 83.22 88.79 84.77 88.44 83.37
Ave 71.01 67.26 67.2 63.3 84.0 79.13 87.29 84.72
Ave 69.13 65.25 81.56 86.00

TABLE II: Sensitivity of the proposed methods and other

methods.
Sensitivity
GC KGC NSGraph Prop. Method Fig. 2: Segmented scans by the proposed FCN for subjects
Subject E.1 E.2 E.1 E.2 E.1 E.2 E.1 E.2 with Intra-retinal.
1 74.39 71.24 83.03 79.4 84.38 81.01 96.55 96.73
2 84.24 80.33 92.42 88.31 94.38 90.36 90.74 93.79
3 50.70 47.26 67.23 62.3 81.82 75.08 86.04 85.23
4 61.70 62.15 79.6 79.88 84.92 84.89 86.91 86.67 FCN to assign them as fluid wrongly. It is shown in Fig. 6 that
5 56.1 53.13 54.73 53.18 67.9 64.26 77.84 73.76 a non-fluid region above RPE is labeled as fluid incorrectly.
6 81.38 76.73 78.8 74.62 87.78 82.34 96.98 96.25 If the neighbors of such regions be darker, they trigger FCN
7 57.14 54.03 56.7 52.05 66.37 63.28 60.5 64.02
to assign them to fluid labels.
8 88.43 87.18 86.39 86.14 92.75 91.56 97.29 96.64
9 79.37 75.32 78.24 74.3 89.31 84.95 96.19 96.1 Undesired artifacts in imaging process such as retinal
10 74.39 70.24 77.86 73.59 82.91 78.44 91.96 91.77 movement, noise, instrument problems and regions of retinal
11 74.29 69.78 77.0 72.63 82.23 77.62 94.51 88.92 without layer structure can create good candidates for fluid
12 62.61 59.6 67.56 64.88 77.15 74.84 91.28 87.45
13 76.83 76.96 84.06 84.02 88.31 88.13 92.38 92.39
regions. A sample of such scans is demonstrated in Fig. 7.
14 62.24 57.92 65.87 61.82 78.27 73.80 80.38 80.64 Elevated RPE creates non-fluid regions referred as drusens.
15 64.41 56.2 66.02 57.76 74.24 66.04 79.39 77.2 These regions are also considered as fluid by FCN. If ROI
16 88.08 84.04 88.63 84.69 90.34 86.29 92.17 86.43 stage segment RPE layers accurately for such cases, this
17 79.57 74.88 89.29 84.13 90.71 85.36 96.56 95.22
18 59.55 51.68 67.45 59.86 81.33 73.3 79.02 77.82 error can be avoided. Also, blood vessels absorbs light waves
19 72.3 67.93 77.73 73.44 83.24 78.57 91.55 87.1 in retinal surface and do not allow to go through retina,
20 75.35 62.68 87.05 74.09 88.61 75.60 86.02 79.85 Therefore, dark regions under blood vessels are appeared
21 77.6 64.02 84.37 69.73 85.61 69.97 94.84 80.33 which is also labeled as fluid incorrectly (see Fig. 8).
22 83.81 73.53 90.66 78.71 87.94 76.8 91.51 79.23
23 85.9 73.21 94.24 79.51 87.77 72.97 96.75 82.09 Finally, as it is depicted in Fig. 9, very small fluid regions
24 74.72 70.33 86.72 82.41 86.85 82.57 88.21 83.04 under RPE are missed by FCN. The main reason is their low
Ave 72.71 67.52 78.4 72.98 83.96 78.25 88.98 85.78 distance to choriod region and inaccurate RPE segmentation.
Ave 70.11 75.69 81.10 87.38
Dark regions between middle layers are assigned as fluid
wrongly. This is because of their similarity with fluid regions.

Authorized licensed use limited to: University of Exeter. Downloaded on June 19,2020 at 23:56:57 UTC from IEEE Xplore. Restrictions apply.
Fig. 3: Segmented scans by the proposed FCN for subjects
with Sub-retinal.
VII. C ONCLUSION
This work presented a fully-automated method for fluid
segmentation based on fully convolutional networks (FCNs)
applied to OCT scans and their corresponding regions of
interest computed by graph shortest path in neutrosophic (NS)
domain. The proposed method was evaluated in 600 OCT
scans and results showed that there are improvements with
respect to dice coefficients and sensitivity in comparison with
two fluid segmentation methods proposed in literature. Future
efforts will be directed to train FCN with augmented training
data by random translation, reflection, rotation, flipping and
cropping to achieve more accurate results. Finally, using more
inputs such as three NS subsets for FCN can be proposed as
another future work.
R EFERENCES
Authorized licensed use limited to: University of Exeter. Downloaded on June 19,2020 at 23:56:57 UTC from IEEE Xplore. Restrictions apply.
Fig. 4: Segmented scans by the proposed FCN for subjects
with Sub-RPE.
[1] A. Rashno, H. SadeghianNejad, A. Heshmati, Highly efficient dimension
reduction for text-independent speaker verification based on relieff
algorithm and support vector machines”, International Journal of Signal
Processing, Image Processing and Pattern Recognition 6 (1) (2013) 91–
108.
[2] A. Rashno, S. M. Ahadi, M. Kelarestaghi, Text-independent speaker
verification with ant colony optimization feature selection and support
vector machine, in: 2015 2nd International Conference on Pattern
Recognition and Image Analysis (IPRIA), IEEE, 2015, pp. 1–5.
[3] A. Rashno, M. Saraee, S. Sadri, Mars image segmentation with most
relevant features among wavelet and color features, in: 2015 AI &
Robotics (IRANOPEN), IEEE, 2015, pp. 1–7.
[4] A. Rashno, F. S. Tabataba, S. Sadri, Image restoration with regularization
convex optimization approach, Journal of Electrical Systems and Signals
2 (2) (2014) 32–36.
[5] A. Rashno, F. S. Tabataba, S. Sadri, Regularization convex optimization
method with l-curve estimation in image restoration, in: 2014 4th
International Conference on Computer and Knowledge Engineering
(ICCKE), IEEE, 2014, pp. 221–226.
[6] Rashno, A., Sadri, S. and SadeghianNejad, H., ”An efficient content-
based image retrieval with ant colony optimization feature selection
schema based on wavelet and color features,” In 2015 The International
Symposium on Artificial Intelligence and Signal Processing (AISP), pp.
59-64, 2015.
[7] Rashno, A. and Rashno, E., ”Content-based image retrieval system with
most relevant features among wavelet and color features,” arXiv preprint
 Fig. 8: Segmentation Error by elevated RPE.

Fig. 5: Segmentation Error by wrongly segmented ROI.

Fig. 9: Segmentation Error by very small fluid regions under
RPE.

arXiv:1902.02059, 2019.
[8] S. Fadaei, A. Rashno, E. Rashno, Content-based image retrieval speedup,
arXiv preprint arXiv:1911.11379.
[9] Heshmati, A., Gholami, M. and Rashno, A., ”Scheme for unsupervised
colour–texture image segmentation using neutrosophic set and non-
subsampled contourlet transform,” IET Image Processing, vol. 10, no.
6, pp. 464-473, 2016.
[10] Rashno, A., Smarandache, F. and Sadri, S., ”Refined neutrosophic sets
in content-based image retrieval application,” In 2017 10th Iranian
Conference on Machine Vision and Image Processing (MVIP), pp. 197-
202, 2017.
[11] Rashno, A. and Sadri, S., ”Content-based image retrieval with color
and texture features in neutrosophic domain,” In 2017 3rd International
Fig. 6: Segmentation Error by non-fluid dark regions. Conference on Pattern Recognition and Image Analysis (IPRIA), pp.
50-55, 2017.
[12] Rashno, E., Akbari, A. and Nasersharif, B., ”A Convoloutional Neural
Network model based on Neutrosophy for Noisy Speech Recognition,”
Infinite Study, 2019.
[13] Rashnoa, E., Minaei-Bidgolia, B. and Guo, Y., ”An effective clustering
method based on data indeterminacy in neutrosophic set domain,”
Infinite Study, 2018.
[14] Rashno, E., Norouzi, S.S., Minaei-Bidgoli, B. and Guo, Y., ”Certainty
of outlier and boundary points processing in data mining,” In 2019 27th
Iranian Conference on Electrical Engineering (ICEE), pp. 1929-1934,
2019.
[15] Rashno, E. and Minaei-Bidgoli, B., ”Boundary points handling for image
edge detection based on Neutrosophic set,” In 2019 5th Conference on
Knowledge Based Engineering and Innovation (KBEI), pp. 886-890,
2019.
[16] A. Rashno, D. D. Koozekanani, P. M. Drayna, B. Nazari, S. Sadri,
H. Rabbani, K. K. Parhi, Fully automated segmentation of fluid/cyst
regions in optical coherence tomography images with diabetic macular
edema using neutrosophic sets and graph algorithms, IEEE Transactions
Fig. 7: Segmentation Error by undesired artifacts. on Biomedical Engineering 65 (5) (2017) 989–1001.
[17] A. Rashno, B. Nazari, D. D. Koozekanani, P. M. Drayna, S. Sadri,
H. Rabbani, K. K. Parhi, Fully-automated segmentation of fluid regions

Authorized licensed use limited to: University of Exeter. Downloaded on June 19,2020 at 23:56:57 UTC from IEEE Xplore. Restrictions apply.
 in exudative age-related macular degeneration subjects: Kernel graph cut
in neutrosophic domain, PloS one 12 (10) (2017) e0186949.
[18] A. Rashno, D. D. Koozekanani, K. K. Parhi, Oct fluid segmentation
using graph shortest path and convolutional neural network, in: 2018
40th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), IEEE, 2018, pp. 3426–3429.
[19] B. Salafian, R. Kafieh, A. Rashno, M. Pourazizi, S. Sadri, Automatic
segmentation of choroid layer in edi oct images using graph theory in
neutrosophic space, arXiv preprint arXiv:1812.01989.
[20] A. Rashno, K. K. Parhi, B. Nazari, S. Sadri, H. Rabbani, P. Drayna,
D. D. Koozekanani, Automated intra-retinal, sub-retinal and sub-rpe
cyst regions segmentation in age-related macular degeneration (amd)
subjects, Investigative Ophthalmology & Visual Science 58 (8) (2017)
397–397.
[21] K. K. Parhi, A. Rashno, B. Nazari, S. Sadri, H. Rabbani, P. Drayna,
D. D. Koozekanani, Automated fluid/cyst segmentation: A quantitative
assessment of diabetic macular edema, Investigative Ophthalmology &
Visual Science 58 (8) (2017) 4633–4633.
[22] J. Kohler, A. Rashno, K. K. Parhi, P. Drayna, S. Radwan, D. D.
Koozekanani, Correlation between initial vision and vision improvement
with automatically calculated retinal cyst volume in treated dme after
resolution, Investigative Ophthalmology & Visual Science 58 (8) (2017)
953–953.
[23] H. Bogunović, F. Venhuizen, S. Klimscha, S. Apostolopoulos, A. Bab-
Hadiashar, U. Bagci, M. F. Beg, L. Bekalo, Q. Chen, C. Ciller, et al.,
Retouch-the retinal oct fluid detection and segmentation benchmark and
challenge, IEEE transactions on medical imaging.
[24] S. H. Kang, H. S. Park, J. Jang, K. Jeon, Deep neural networks for the
detection and segmentation of the retinal fluid in oct images, in: Proc.
MICCAI Retinal OCT Fluid Challenge (RETOUCH), 2017, pp. 9–14.
[25] A. G. Roy, S. Conjeti, S. P. K. Karri, D. Sheet, A. Katouzian,
C. Wachinger, N. Navab, Relaynet: retinal layer and fluid segmentation
of macular optical coherence tomography using fully convolutional
networks, Biomedical optics express 8 (8) (2017) 3627–3642.
[26] F. G. Venhuizen, B. van Ginneken, B. Liefers, M. J. van Grinsven,
S. Fauser, C. Hoyng, T. Theelen, C. I. Sánchez, Robust total retina
thickness segmentation in optical coherence tomography images using
convolutional neural networks, Biomedical optics express 8 (7) (2017)
3292–3316.
[27] T. Schlegl, S. M. Waldstein, H. Bogunovic, F. Endstraßer,
A. Sadeghipour, A.-M. Philip, D. Podkowinski, B. S. Gerendas,
G. Langs, U. Schmidt-Erfurth, Fully automated detection and
quantification of macular fluid in oct using deep learning,
Ophthalmology 125 (4) (2018) 549–558.
[28] F. G. Venhuizen, B. van Ginneken, B. Liefers, F. van Asten, V. Schreur,
S. Fauser, C. Hoyng, T. Theelen, C. I. Sánchez, Deep learning approach
for the detection and quantification of intraretinal cystoid fluid in
multivendor optical coherence tomography, Biomedical optics express
9 (4) (2018) 1545–1569.
[29] A. Rashno, S. Sadri, H. SadeghianNejad, An efficient content-based im-
age retrieval with ant colony optimization feature selection schema based
on wavelet and color features, in: 2015 The International Symposium
on Artificial Intelligence and Signal Processing (AISP), IEEE, 2015, pp.
59–64.
[30] A. Rashno, B. Nazari, S. Sadri, M. Saraee, Effective pixel classification
of mars images based on ant colony optimization feature selection and
extreme learning machine, Neurocomputing 226 (2017) 66–79.
[31] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for
semantic segmentation, in: Proceedings of the IEEE conference on
computer vision and pattern recognition, 2015, pp. 3431–3440.
[32] Boykov Y, Funka-Lea G. “Graph cuts and efficient nd image segmen-
tation,” International journal of computer vision, vol. 70, no. 2, pp.
109–131, 2006.
[33] Salah MB, Mitiche A, Ayed IB. “Multiregion image segmentation by
parametric kernel graph cuts,” IEEE Transactions on Image Processing,
vol. 20, no. 2, pp. 545–557, 2011.

Authorized licensed use limited to: University of Exeter. Downloaded on June 19,2020 at 23:56:57 UTC from IEEE Xplore. Restrictions apply.