Retinal Image Segmentation with Small Datasets

Nchongmaje Ndipenoch, Alina Miron, Zidong Wang and Yongmin Li

Department of Computer Science, Brunel University London, Uxbridge, UB8 3PH, U.K.

Keywords:

Medical Imaging, Retinal Layers and Fluid Segmentation, Deep Learning, Convolutional Neural Network,

Optical Coherence Tomography (OCT).

Abstract:

Many eye diseases like Diabetic Macular Edema (DME), Age-related Macular Degeneration (AMD), and

Glaucoma manifest in the retina, can cause irreversible blindness or severely impair the central version. The

Optical Coherence Tomography (OCT), a 3D scan of the retina with high qualitative information about the

retinal morphology, can be used to diagnose and monitor changes in the retinal anatomy. Many Deep Learning

(DL) methods have shared the success of developing an automated tool to monitor pathological changes in the

retina. However, the success of these methods depends mainly on large datasets. To address the challenge from

very small and limited datasets, we proposed a DL architecture termed CoNet (Coherent Network) for joint

segmentation of layers and ﬂuids in retinal OCT images on very small datasets (less than a hundred training

samples). The proposed model was evaluated on the publicly available Duke DME dataset consisting of 110 B-

Scans from 10 patients suffering from DME. Experimental results show that the proposed model outperformed

both the human experts’ annotation and the current state-of-the-art architectures by a clear margin with a mean

Dice Score of 88% when trained on 55 images without any data augmentation.

1 INTRODUCTION

Diabetic retinopathy (DR), a disease that damages the

blood vessels in the retina, is the most common cause

of blindness among working-aged adults in the United

States (Kles and Anderson, 2007). Among those af-

fected, approximately 21 million people develop dia-

betic macular edema (DME) (Bresnick, 1986). DME

is the accumulation of ﬂuid in the macula that can

damage the blood vessels in the eye due to high blood

sugar over time. The macula is the retina’s centre at

the back of the eye, where vision is the sharpest.

Presently an effective treatment of eye diseases

exists in the form of anti-vascular endothelial growth

factor (anti-VEGF) therapy (Shienbaum et al., 2013).

However, the effectiveness of the treatment depends

on early diagnosis and frequent monitoring of the

progress of the disease. Also, anti-VEGF drugs are

expensive and need to be administered regularly.

Early diagnosis, effective frequent monitoring,

and behavioural advice from ophthalmologists, such

as diets and regular exercises, are key factors in pre-

venting or slowing down the disease’s progress. Still,

as of today, these are mostly done manually, which

is time-consuming, laborious, and prone to errors.

Hence, there is the need to develop an automated tool

to monitor retinal morphology and ﬂuid accumulation

properly.

The Optical Coherence Tomography (OCT), a

high-resolution 3D non-invasive imaging modality of

the retina acquiring a series of cross-sectional slices

(B-scans), provides qualitative information and visu-

alisations of the retinal structure. The development

of an automated method to study the retina anatomy

from OCT B-Scans and hence the evaluation of eye

pathogens like DME will be of high value and impor-

tance.

To address the above problem, we propose a deep

learning based model, termed CoNet (Coherent Net-

work), for simultaneously segmenting layers and ﬂuid

in retinal OCT B-Scans. In contrast to the common

approach of treating retinal layers and ﬂuid regions

separately, CoNet provides an automatic solution for

simultaneously segmenting both.

The rest of the paper is organized as follows. A

brief review of the previous studies is provided in Sec-

tion 2. The description of the proposed method is

put forward in Section 3. The experiments and re-

sult analysis are presented in Section 4. Finally, the

conclusion with our contributions is described in Sec-

tion 5.

Ndipenoch, N., Miron, A., Wang, Z. and Li, Y.

Retinal Image Segmentation with Small Datasets.

DOI: 10.5220/0011779200003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 2: BIOIMAGING, pages 129-137

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

129

2 BACKGROUND

The OCT was developed in the 1990s by (Huang

et al., 1991) but only became commercially available

in 2006. It permits fast image acquisition and suc-

cess in quantitative analysis because of its high qual-

ity and resolution. Some of the earliest segmenta-

tion approaches of retinal images include: segmen-

tation of retinal layers in OCT images using the graph

method (Garvin et al., 2009), segmentation of ﬂuid in

the retina in patients suffering from Macular Edema

(ME) by (Abr

amoff et al., 2010), and the segmen-

tation of ﬂuid using the active contours approach by

(Fernandez, 2005).

Other approaches used for the segmentation of

retinal OCT include: The traditional graph-cut meth-

ods by (Salazar-Gonzalez et al., 2014; Salazar-

Gonzalez et al., 2010; Salazar-Gonzalez et al., 2011;

Kaba et al., 2015), the Markov Random Fields by

(Salazar-Gonzalez et al., 2012; Wang et al., 2017),

probabilistic modelling (Kaba et al., 2014; Kaba et al.,

2013), dynamic programming by (Chiu et al., 2015),

level set by (Dodo et al., 2019a; Dodo et al., 2019b)

and a combination of a fuzzy C-means and level set

contour by (Wang et al., 2016).

Recent approaches have shifted to the Deep

Learning methods, some of which will be reviewed

brieﬂy below.

(Fang et al., 2017) presented an approach to seg-

ment nine retinal boundaries from retinal OCT us-

ing a combination of convolutional neural network

(CNN) and graph search. Their approach was tested

on 60 volumes consisting of 2915 B-scans from 20

human eyes suffering with dry AMD. (Fauw et al.,

2018) presented a 3D U-Net (Cicek et al., 2016)

model framework for diagnosis and referral in retinal

disease. Their dataset consisted of 14,884 volumes

from 7,621 patients. The ReLayNet was presented by

(Roy et al., 2017), which is a 2D-like U-Net archi-

tecture to segment layers and ﬂuids in the OCT im-

ages. The method was validated on the Duke DME

dataset (Chiu et al., 2015) which consists of 110 B-

Scans from 10 patients suffering from Diabetic Mac-

ular Edema (DME). Another CNN approach was re-

ported by (Lee et al., 2017) to segment ﬂuid from

1,289 OCT images from patients suffering from Mac-

ular Edema (ME). (Lu et al., 2017) reported a CNN

approach to detect and segment three retinal ﬂuid

types from OCT images and their method was val-

idated on the RETOUCH dataset (Bogunovi

c et al.,

2019). A neutrosophic transformation and a graph-

based shortest path to segment ﬂuid in OCT images

was presented in (Rashno et al., 2017). Their method

was also evaluated on the DME dataset. A Deep

Learning approach for simultaneous segmentation of

layers and ﬂuids in retinal OCT B-Scans from patients

suffering from AMD is proposed in (Ndipenoch et al.,

2022). The algorithm consists of the traditional U-Net

with an encoding and a decoding path, skip connec-

tion blocks, squeeze and exiting blocks and an Atrous

Spatial Pyramid Pooling (ASPP) block. The method

is validated on 1136 B-Scans from 24 patients. Other

CNN approaches to segment ﬂuids in retinal OCT

modality includes (Schlegl et al., 2015; Venhuizen

et al., 2018; Gopinath and Sivaswamy, 2018; Girish

et al., 2018).

Previous studies indicate that in this domain U-

Net and CNN are the most popular methods used

but, U-Net tends to outperform CNN, and hence U-

Net is the preferred choice in many applications. An

overview of the related work is summarised in Table 1

below.

Table 1: Overview of the related work with references and

corresponding ﬂuid and disease types.

Reference Class Disease

(Fernandez, 2005) Fluid AMD

(Garvin et al., 2009) Fluid –

(Abr

amoff et al., 2010) Fluid ME

(Salazar-Gonzalez et al., 2010) – –

(Salazar-Gonzalez et al., 2012) Optic disc DR

(Salazar-Gonzalez et al., 2014) B. vessel DR

(Chiu et al., 2015) Fluid DME

(Wang et al., 2016) Fluid DME

(Wang et al., 2017) Fluid DME

(Fang et al., 2017) Layers AMD

(Dodo et al., 2019a) Layers –

(Loo et al., 2018) Fluid ME

(Schlegl et al., 2015) Fluid –

(Venhuizen et al., 2018) Fluid AMD

(Gopinath and Sivaswamy, 2018)Fluid DME

(Girish et al., 2018) Fluid ME

(Lu et al., 2017) Fluid ME

(Fauw et al., 2018) Fluid –

(Roy et al., 2017) Fluid DME

(Rashno et al., 2017) Fluid DME

(Ndipenoch et al., 2022) Fluid/LayerAMD

3 METHOD

Deep Learning methods have had success in image

segmentation (pixel-wise classiﬁcation) but this de-

pends hugely on large datasets. In medical imaging

obtaining a dataset is very challenging and often very

small and limited. We aim to provide a model that

performs very well on very small and limited datasets

BIOIMAGING 2023 - 10th International Conference on Bioimaging

130

of less than a hundred training images.

The proposed CoNet model is based on commonly

used U-Net architecture (Ronneberger et al., 2015)

but adapted to the speciﬁc problem of retina image

segmentation on very small datasets. The model

architecture consists of an encoding path, a decod-

ing path, a bottleneck, a classiﬁcation layer, and an

Atrous Spatial Pyramid Pooling (ASPP) block, as

shown in Figure 1. In this section, we will explain our

method, the changes we made and how it is different

to the U-Net.

3.1 Encoding Path

The encoding path is use to capture local contextual

and spatial information. As we move down the en-

coding or contracting path the feature map is reduced

by half after every convolutional block by a convolu-

tional operation at the downsampling layer. A total of

three convolutional blocks are used and for each block

the convolutional operations are set up in the order of:

(1) convolutional layer which converts all the pixels of

the receptive ﬁeld into a single value and passes it to

the next operation, (2) ReLU activation, to circumvent

the problem of vanishing gradient, (3) batch normali-

sation layer to prevent over-ﬁtting during training, (4)

convolutional layer, and (5) ReLU activation. For the

convolutional operations a rectangular kernel size of

9 × 3 is used to match the rectangular shape of the

original B-Scans as opposed to the square kernel size

of 3 × 3 used in U-Net, also to ensure that the fea-

ture map before and after the convolutional layer is

the same padding was set to 3 × 1 still to match the

rectangular shape of the original B-Scans and ﬁnally

to ensure no overlapping when constructing the fea-

ture map a stride of 1 was used.

3.2 Decoding Path

The decoding path is used to enable precise localiza-

tion of the pixel and as we move up the decoding or

expansive path, before each convolutional block the

size of the feature maps is double by a convolutional

operation at the upsmapling layer. Same as in the en-

coding path a total of three convolutional blocks are

used and set up in the same order as mentioned in 3.1.

In addition to that, the upsampling layer was used to

to double the size of the feature map by capturing spa-

tial information from the previous feature map and

also to ensure that the size of the input image is the

same as the output image, the concatenating layer is

used to concatenate images from the encoder phase to

their corresponding decoder phase.

Because of the very small size of the dataset (only

55 B-Scans for training), we reduce the depth of the

network from 5 convolutional blocks as in the stan-

dard 2D U-Net to 3 in both the encoding and decoding

phase. Furthermore reducing the depth of the network

trains the model faster and uses less memory because

less parameters are used.

3.3 Bottleneck

Between the encoding and decoding paths is a bottle-

neck. The bottleneck serves as a bridge layer between

the encoding or contracting path and the decoding or

expansive path to ensure a smooth transition from one

path to the other. In CoNet, the bottleneck is made up

of a convolutional block that consists of six parts or

layers in the order of convolutional layer, ReLU acti-

vation, batch normalization layer, convolutional layer,

and ReLU activation. These layers were used for the

same reasons as mentioned in section 3.1

3.4 Atrous Spatial Pyramid Pooling

(ASPP)

The ASSP is a technique used to capture global con-

textual information on a multi-scale by applying mul-

tiple parallel ﬁlters with different frequencies or di-

lating rates on a given image or feature map (Chen

et al., 2017). To enhance the performance of the

ASSP block, global average pooling is used at the

last feature map to further capture global information.

The output of the parallel ﬁlters are concatenated us-

ing a 1 × 1 convolution to get the ﬁnal results. While

ASSP is designed to capture global information, it is

also computational efﬁcient.

No ASPP block is used in the standard 2D U-Net.

We have used an ASPP block as the input layer of

CoNet, and it consists of 4 parallel ﬁlters with a dilat-

ing rate of 6, 12, 18, and 24. To circumvent the prob-

lem of high ﬂuid variability (the ﬂuid class was absent

in some B-Scans for some patients) and an imbalance

dataset we have used the ASPP block in CoNet. The

ASPP block used in CoNet is illustrated in Figure 2.

3.5 Classiﬁcation Layer

At the classiﬁcation layer, we have used a convolu-

tional layer with a kernel size of 3, stride of 1 and

padding of 1. The task is to determine which class

out of the ten labelled classes each voxel or pixel of

the ﬁnal feature map is assigned to. In the 2D U-Net

the same goal was achieved using the SoftMax acti-

vation.

Retinal Image Segmentation with Small Datasets

131

Figure 1: Architecture of the proposed CoNet, consisting of the ASPP block, the encoding path, the bottleneck, the decoding

path and the classiﬁcation layer.

Figure 2: The ASPP blocks in the proposed model capture multi-scale information by applying multiple parallel ﬁlters with

different frequencies.

4 EXPERIMENTS

4.1 Dataset

The dataset used in the experiments is the Duke DME

dataset (Chiu et al., 2015) which is publicly avail-

able. It is consists of 110 B-scans from 10 patients

with severe DME pathology. It was collected using

the standard Spectralis (Heidelberg Engineering, Hei-

delberg, Germany). The volumetric scans were Q =

61 B-scans N = 768 A-scans with an axial resolu-

tion of 3.87 m/pixel, lateral resolution ranging from

11.07 - 11.59 m/pixel, and azimuthal resolution rang-

ing from 118 - 128 m/pixel. Annotation of the im-

ages was done by 2 human experts for three categories

(layer, ﬂuid and background) consisting of 10 classes

(1 ﬂuid, 2 backgrounds and 7 layers). In the past the

retinal OCT is layered for 10 layers but for lucidity

they are grouped into 7 distinct classes which are: In-

ner Limiting Membrane (ILM), Nerve Fiber ending

to Inner Plexiform Layer (NFL-IPL), Inner Nuclear

Layer (INL), and the Outer plexiform Layer (OPL).

A ﬂuid class was identiﬁed and the two background

classes were the area above the retinal and the area

below the retinal.

In this work the annotated colors for the classes

are: Black which is the area above and below the reti-

nal, light green which is the ILM layer, yellow which

is the area between NFL and IPL layers, Blue which

is the INL, Pink which is the OPL layer, light blue

BIOIMAGING 2023 - 10th International Conference on Bioimaging

132

Figure 3: Annotation and labelling of the 10 segments (7 retinal layers, 2 backgrounds and 1 ﬂuid) in the Duke DME dataset.

which is the area between the ONL and ISM layers,

Green which is the ISE layer, White which is the RPE,

and Red which is the ﬂuid. An example of annotation

and labelling of classes is shown in Figure 3.

It is worth to note that the Duke DME dataset was

collected for two problems (layer and ﬂuid segmenta-

tion). Also to add to the complexity of the dataset, the

ﬂuid class demonstrates a high level of variability and

was not present in some B-Scans for some patients.

4.2 Training and Testing

Retinal OCT layers are complex in nature, coupling

with high level of variability of the ﬂuid classes. It

is therefor a common practice to do segmentation of

layers and detection of ﬂuids separately, but in this

work we performed both simultaneously which is a

harder task.

In this work training and testing were done using

annotation from expert 2.

Training was done on 55 B-Scans and no data aug-

mentation was used. We used B-Scans instead of the

entire volumes because of the anisotropic resolution

of OCT volumes and the present of possible motion

artifacts across B-scans.

K-fold cross validation was used for training, vali-

dation and testing. Parameters and environmental set-

tings where the same for the proposed model and the

comparison models to ensure fairness. B-Scans from

5 patients were used per fold, that is patients 1-5 in

the ﬁrst fold and 6-10 in the second. To eliminate

bias the use of adjacent B-Scans in training, valida-

tion, and testing is not recommended. Across all the

experiments the parameters were set up as, still, the

same as in the comparison models: the value of k

was 2, the original B-Scans were resized to 512 ×512

pixels, the loss function used was Categorical cross-

entropy which provides an estimated probability be-

tween the predicted voxels and the ground truth for

the current state of the model, the batch size was set

to 4, the cost function was optimized using AdaDelta

and back-propagation using the chain rule, by default

the learning rate is set by AdaDelta as explained in

(Zeiler, 2012), and the model was trained for 200

epochs. The AdaDelta’s equation is shown in Eqn (1).

∆θ

= −

E[g

]

+ ε

(1)

Dice score also known as the F1 Score was the

evaluation metric used to measure the performance of

the algorithm. It gives a score of how well the pixels

are classiﬁed to belong to the correct class per class

in the range from 0 to 1 with 0 being the worst and

1 the perfect classiﬁcation. In many medical image

segmentation problems Dice Score is the preferred

choice. The formula to calculate the Dice score is

shown in Eqn (2).

DSC =

2|X ∩Y |

|X| + |Y |

(2)

The ﬂuid class was missing for some B-Scans for

some patients. Hence during testing the calculation of

Dice score for the ﬂuid class was exempted from B-

scans with no ﬂuid reference for that patient to avoid

over estimation or under estimation.

The models were trained on a GPU work station

with NVIDIA RTX A6000 48GB. The models were

implemented in Python, using PyTorch library.

4.3 Results

In this section we present and analyse the segment

class results measured in Dice score, of the proposed

CoNet and compare them to the comparison models

(the state-of-the-art, ReLayNet and baselines U-Net)

and the human expert annotation (Inter-observer).

Retinal Image Segmentation with Small Datasets

133

Figure 4: A Bar chart comparison of the performance in Dice score grouped by segment class of the inter-observers, U-Net,

RelayNet and the proposed CoNet.

A bar chart of the segmentation grouped by seg-

ment classes is shown in Figure 4, the Dice Scores

in Table 2, examples of the visualization results to-

gether with their annotations are illustrated in Fig-

ure 5 and a zoom in of a visualization output example

from CoNet is illustrated in Figure 6. Orange arrows

are used to show ﬁne details in the annotated B-Scans

that were picked up by corresponding models. Anal-

ysis from our results show that:

1. The proposed model CoNet outperforms the hu-

man experts, the baseline (U-Net) and current

state-of-the-art model ReLayNet in every single

class by a clear margin.

2. CoNet obtained a Dice Score of 77% which

is 19% greater than the human experts’ (inter-

observer) in the ﬂuid class which was the most

difﬁcult to segment.

3. We obtained a Dice score of 90% and above in 8

out of the 10 classes.

4. The baseline U-Net, the state-of-the-art architec-

ture RelayNet, and the proposed CoNet all ob-

tained a perfect Dice Score of 100% in both back-

ground classes (area above and below the retinal).

5. CoNet obtained an overall mean Dice Score of

88% which is 8% higher than that of the human

experts’ annotation results of 80%.

6. We noticed an increase of performance from the

standard U-Net to a shallower and less com-

plex architectures in the order of ReLayNet, and

CoNet.

Table 2: Segmentation performance (Dice Scores) by seg-

ment classes (rows) and models (columns).

Inter Obs. U-Net ReLayNet

Proposed

Fluid 0.58 0.70 0.75 0.77

NFL 0.86 0.85 0.88 0.90

GCL IPL 0.89 0.90 0.92 0.93

INL 0.77 0.77 0.82 0.83

OPL 0.72 0.74 0.80 0.82

ONL ISM 0.87 0.88 0.91 0.93

ISE 0.85 0.86 0.92 0.93

OS RPE 0.82 0.84 0.89 0.91

5 CONCLUSIONS

In this paper, we have presented the CoNet, a Deep

Learning approach for the joint layers and ﬂuids seg-

mentation of retinal OCT B-Scans. The model was

evaluated on the publicly available Duke DME dataset

consisting of 110 B-Scans without any data augmen-

tation. Taking into consideration of the speciﬁc char-

acteristics of the problem, in particular with small

available dataset, the proposed model has the follow-

ing distinct features compared to the previous U-Net

based models:

1. We have reduced the depth of the network from 5

BIOIMAGING 2023 - 10th International Conference on Bioimaging

134

Figure 5: Examples to illustrate the visualisation output of U-Net, ReLayNet and the proposed CoNet, in order of the inputs,

annotations and outputs with orange arrows to demonstrate ﬁne details picked up by the models.

Figure 6: A Zoom in of the B-scan of the second row in Figure 5 to highlight the ﬁne details picked up by the CoNet using

orange arrows.

to 3 convolutional blocks. This was done because

of the very small size of the dataset (only 55 B-

Scans were used during training and the rest 55 for

testing). Deeper and more complex architectures

turn to yield poorer results. Also, reducing the

depth of the network enhances the training speed

of the network and uses less memory since less

parameters are used.

2. We have introduced an ASPP block at the input

layer to capture global information from the input

image, because the dataset demonstrates a high

level of variability. The ﬂuid class was not present

in some B-Scans for some patients.

3. At the classiﬁcation layer to classify each pixel to

belong to one of the 10 classes we have used a

convolutional layer instead of the SoftMax acti-

vation. This was because using the convolutional

layer for the classiﬁcation of the ﬂuid class which

is highly variable yielded a better and more accu-

rate results.

4. We have used a rectangular kernel size of 9 × 3

instead of the square kernel size of 3 × 3 to match

Retinal Image Segmentation with Small Datasets

135

the rectangular shape of the original B-Scans.

Evaluation was done on the basis of Dice Score

which is a standard method of evaluating segmenta-

tion problems. Experimental results show that the

proposed model outperformed both the human ex-

perts’ annotation and the current state-of-the-art ar-

chitectures by a clear margin, even on a very small,

imbalanced and complex dataset with a high degree

of presence of pathology that severely affects the nor-

mal morphology of the retina.

The dataset was collected for 2 problems (layers

and ﬂuid segmentation) which can be experimented

separately but we decided to do both jointly together

which is a more challenging task.

The CoNet can be directly applied to solve real

world problems and to monitor the progress of eye

diseases such as diabetic macular edema (DME), age-

related macular degeneration (AMD) and Glaucoma.

In the future we will evaluate the CoNet on other

benchmark datasets, and compare our results to other

state-of-the-art models. Also we plan to extend the

current 2D network to 3D.

REFERENCES

Abr

amoff, M. D., Garvin, M. K., and Sonka, M. (2010).

Retinal imaging and image analysis. IEEE reviews in

biomedical engineering, 3:169–208.

Bogunovi

c, H., Venhuizen, F., Klimscha, S., Apostolopou-

los, S., Bab-Hadiashar, A., Bagci, U., Beg, M. F.,

Bekalo, L., Chen, Q., Ciller, C., et al. (2019). Re-

touch: the retinal oct ﬂuid detection and segmentation

benchmark and challenge. IEEE transactions on med-

ical imaging, 38(8):1858–1874.

Bresnick, G. H. (1986). Diabetic macular edema: a review.

Ophthalmology, 93(7):989–997.

Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and

Yuille, A. L. (2017). Deeplab: Semantic image seg-

mentation with deep convolutional nets, atrous convo-

lution, and fully connected crfs. IEEE transactions on

pattern analysis and machine intelligence, 40(4):834–

848.

Chiu, S., Allingham, M., Mettu, P., Cousins, S., Izatt, J., and

Farsiu, S. (2015). Kernel regression based segmen-

tation of optical coherence tomography images with

diabetic macular edema. Biomedical Optics Express,

Cicek, O., Abdulkadir, A., Lienkamp, S., Brox, T., and Ron-

neberger, O. (2016). 3D U-Net: Learning dense volu-

metric segmentation from sparse annotation.

Dodo, B. I., Li, Y., Kaba, D., and Liu, X. (2019a). Retinal

layer segmentation in optical coherence tomography

images. IEEE Access, 7:152388–152398.

Dodo, B. I., Li, Y., and Liu, X. (2019b). Level set segmenta-

tion of retinal oct images. In International Conference

on Bioimaging. Czech Republic.

Fang, L., Cunefare, D., Wang, C., Guymer, R. H., Li, S.,

and Farsiu, S. (2017). Automatic segmentation of

nine retinal layer boundaries in oct images of non-

exudative amd patients using deep learning and graph

search. Biomedical optics express, 8(5):2732–2744.

Fauw, J., Ledsam, J., Romera-Paredes, B., Nikolov, S.,

Tomasev, N., Blackwell, S., Askham, H., Glorot, X.,

ODonoghue, B., Visentin, D., Driessche, G., Lak-

shminarayanan, B., Meyer, C., Mackinder, F., Bou-

ton, S., Ayoub, K., Chopra, R., King, D., Karthike-

salingam, A., and Ronneberger, O. (2018). Clinically

applicable deep learning for diagnosis and referral in

retinal disease. Nature Medicine, 24.

Fernandez, D. C. (2005). Delineating ﬂuid-ﬁlled region

boundaries in optical coherence tomography images

of the retina. IEEE transactions on medical imaging,

24(8):929–945.

Garvin, M. K., Abramoff, M. D., Wu, X., Russell, S. R.,

Burns, T. L., and Sonka, M. (2009). Automated 3-

d intraretinal layer segmentation of macular spectral-

domain optical coherence tomography images. IEEE

transactions on medical imaging, 28(9):1436–1447.

Girish, G., Thakur, B., Chowdhury, S. R., Kothari, A. R.,

and Rajan, J. (2018). Segmentation of intra-retinal

cysts from optical coherence tomography images

using a fully convolutional neural network model.

IEEE journal of biomedical and health informatics,

23(1):296–304.

Gopinath, K. and Sivaswamy, J. (2018). Segmentation of

retinal cysts from optical coherence tomography vol-

umes via selective enhancement. IEEE journal of

biomedical and health informatics, 23(1):273–282.

Huang, D., Swanson, E. A., Lin, C. P., Schuman, J. S., Stin-

son, W. G., Chang, W., Hee, M. R., Flotte, T., Gregory,

K., Puliaﬁto, C. A., et al. (1991). Optical coherence

tomography. science, 254(5035):1178–1181.

Kaba, D., Salazar-Gonzalez, A. G., Li, Y., Liu, X., and

Serag, A. (2013). Segmentation of retinal blood ves-

sels using gaussian mixture models and expectation

maximisation. In International Conference on Health

Information Science, pages 105–112.

Kaba, D., Wang, C., Li, Y., Salazar-Gonzalez, A., Liu, X.,

and Serag, A. (2014). Retinal blood vessels extrac-

tion using probabilistic modelling. Health Informa-

tion Science and Systems, 2(1):2.

Kaba, D., Wang, Y., Wang, C., Liu, X., Zhu, H., Salazar-

Gonzalez, A. G., and Li., Y. (2015). Retina layer

segmentation using kernel graph cuts and continuous

max-ﬂow. Optics Express, 23(6):7366–7384.

Kles, J. B. M.-K. A. and Anderson, P. W. (2007). How the

diabetic eye loses vision.

Lee, C. S., Tyring, A. J., Deruyter, N. P., Wu, Y., Rokem,

A., and Lee, A. Y. (2017). Deep-learning based, au-

tomated segmentation of macular edema in optical

coherence tomography. Biomedical optics express,

8(7):3440–3448.

Loo, J., Fang, L., Cunefare, D., Jaffe, G. J., and Farsiu, S.

(2018). Deep longitudinal transfer learning-based au-

tomatic segmentation of photoreceptor ellipsoid zone

defects on optical coherence tomography images of

BIOIMAGING 2023 - 10th International Conference on Bioimaging

136

macular telangiectasia type 2. Biomedical optics ex-

press, 9(6):2681–2698.

Lu, D., Heisler, M., Lee, S., Ding, G., Sarunic, M. V., and

Beg, M. F. (2017). Retinal ﬂuid segmentation and de-

tection in optical coherence tomography images us-

ing fully convolutional neural network. arXiv preprint

arXiv:1710.04778.

Ndipenoch, N., Miron, A., Wang, Z., and Li, Y. (2022). Si-

multaneous segmentation of layers and ﬂuids in retinal

oct images. In 2022 15th International Congress on

Image and Signal Processing, BioMedical Engineer-

ing and Informatics (CISP-BMEI), pages 1–6. IEEE.

Rashno, A., Koozekanani, D. D., Drayna, P. M., Nazari, B.,

Sadri, S., Rabbani, H., and Parhi, K. K. (2017). Fully

automated segmentation of ﬂuid/cyst regions in opti-

cal coherence tomography images with diabetic mac-

ular edema using neutrosophic sets and graph algo-

rithms. IEEE Transactions on Biomedical Engineer-

ing, 65(5):989–1001.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Roy, A. G., Conjeti, S., Karri, S. P. K., Sheet, D., Katouzian,

A., Wachinger, C., and Navab, N. (2017). Relaynet:

retinal layer and ﬂuid segmentation of macular optical

coherence tomography using fully convolutional net-

works. Biomedical optics express, 8(8):3627–3642.

Salazar-Gonzalez, A., Kaba, D., Li, Y., and Liu, X. (2014).

Segmentation of the blood vessels and optic disk in

retinal images. IEEE journal of biomedical and health

informatics, 18(6):1874–1886.

Salazar-Gonzalez, A., Li, Y., and Kaba, D. (2012). MRF

reconstruction of retinal images for the optic disc seg-

mentation. In International Conference on Health In-

formation Science, pages 88–99.

Salazar-Gonzalez, A. G., Li, Y., and Liu, X. (2010). Retinal

blood vessel segmentation via graph cut. In Interna-

tional Conference on Control Automation Robotics &

Vision, pages 225–230.

Salazar-Gonzalez, A. G., Li, Y., and Liu, X. (2011). Op-

tic disc segmentation by incorporating blood vessel

compensation. In IEEE Third International Workshop

On Computational Intelligence In Medical Imaging,

pages 1–8.

Schlegl, T., Waldstein, S. M., Vogl, W.-D., Schmidt-Erfurth,

U., and Langs, G. (2015). Predicting semantic de-

scriptions from medical images with convolutional

neural networks. In International Conference on In-

formation Processing in Medical Imaging, pages 437–

448. Springer.

Shienbaum, G., Garcia Filho, C. A. A., Flynn Jr, H. W.,

Nunes, R. P., Smiddy, W. E., and Rosenfeld, P. J.

(2013). Management of submacular hemorrhage sec-

ondary to neovascular age-related macular degen-

eration with anti–vascular endothelial growth factor

monotherapy. American journal of ophthalmology,

155(6):1009–1013.

Venhuizen, F. G., van Ginneken, B., Liefers, B., van As-

ten, F., Schreur, V., Fauser, S., Hoyng, C., Theelen,

T., and S

anchez, C. I. (2018). Deep learning approach

for the detection and quantiﬁcation of intraretinal cys-

toid ﬂuid in multivendor optical coherence tomogra-

phy. Biomedical optics express, 9(4):1545–1569.

Wang, C., Wang, Y., and Li, Y. (2017). Automatic choroidal

layer segmentation using markov random ﬁeld and

level set method. IEEE journal of biomedical and

health informatics.

Wang, J., Zhang, M., Pechauer, A. D., Liu, L., Hwang,

T. S., Wilson, D. J., Li, D., and Jia, Y. (2016). Au-

tomated volumetric segmentation of retinal ﬂuid on

optical coherence tomography. Biomedical optics ex-

press, 7(4):1577–1589.

Zeiler, M. D. (2012). Adadelta: an adaptive learning rate

method. arXiv preprint arXiv:1212.5701.

Retinal Image Segmentation with Small Datasets

137