A Deep Learning Approach for the Detection and Identification of Neovascularization in Fundus Images

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 06 | Jun 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 914
A Deep Learning Approach for the Detection and Identification of
Neovascularization in Fundus Images
Sheik Arshad1, Rahul Paul2, Birali Prasanthi3, K Mahesh Kumar4
1 Student, Bachelors in CSE, Mahatma Gandhi Institute of Technology, Hyderabad, Telangana, India
2 Student, Bachelors in CSE, Mahatma Gandhi Institute of Technology, Hyderabad, Telangana, India
3Assistant Professor, Dept. of Computer Science and Engineering, Mahatma Gandhi Institute of Technology,
Hyderabad, Telangana, India
4Associate Professor, Dept. of Computer Science and Engineering, Mahatma Gandhi Institute of Technology,
Hyderabad, Telangana, India
----------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Diabetic patients are at ahighriskof developing
a retinal disorder called Proliferative Diabetic Retinopathy
(PDR). In PDR Neovascularization is considered as one of the
major conditions in which there is an abnormal random
growth of blood vessels on the retina. Neovascularization can
cause severe vision loss and blindness if it is not detected and
treated in its early stages. Fundus images include the images
of the rear of an eye. Using these fundus images
Neovascularization can be detected and classified intoseveral
stages. Neovascularization has a small size and random
abnormal growth pattern. This could be a challenging task to
detect with normal Image processing techniques. Deep
learning methods can be used to detect Neovascularization
because of their ability to perform automatic feature
extraction on objects with complex features. The proposed
system is implemented based on the performance of popular
pre-trained deep neural networks suchasInceptionResNetV2,
DenseNet, ResNet50, ResNet18, AlexNet, andVGG19networks.
The best Convolutional Neural Network (CNN) model can be
used to build and implement the Neovascularizationdetection
and classification Model.
Key Words: CNN, Deep Learning, Neovascularization,
Fundus Images, Pre-trained Networks.
1. INTRODUCTION
Neovascularization is the formation of abnormal blood
vessels in the retina, which can be a complication of various
retinal diseases, including diabetic retinopathy. The
presence of neovascularization can lead toseverevision loss
and blindness if left untreated. Thus, early detection and
timely intervention are crucial for preventing or minimizing
the damage caused bythiscondition. TheNeovascularization
is further classified into 5 conditions - Healthy Eye, Mild,
Moderate, Proliferate, and Severe.
Fundus images correspond to the retinal view of an eye
which represents the rear of aneye. Theseimagesaremostly
used for detecting eye-related disorders. Fundus images are
commonly used for the screening and diagnosis of diabetic
retinopathy. However, detecting neovascularizationinthese
images can be challenging due to the presence of various
confounding factors such as image noise,variabilityinimage
quality, and the presence of other retinal abnormalities.
Hence, developing automated and accurate methods for
detecting neovascularization in fundus images is essential.
Deep learning techniques, particularly convolutional neural
networks (CNNs),haveshownremarkablesuccessinvarious
image-processing tasks, including medical image analysis.
These techniques can learn complex features from the input
images and detect the presence of Neovascularization and
also classify them into different categories(Mild, Moderate,
Healthy, Proliferate, Severe).
The proposed approach for neovascularization detection in
fundus imagescombinesdeeplearningandimageprocessing
techniques to develop a reliable and accurate system. By
using a minimum amount of fundus image dataset, the deep
learning model can learn to identify and classify the
condition of neovascularizationinfundusimageswitha finer
accuracy. Overall, the proposed approach can help improve
the early detection and management of neovascularization,
leading to better patient outcomes and reduced healthcare
costs.
1.1 LITERATURE SURVEY
Several research studies have proposed divergent image
processing methods to detect Neovascularization in fundus
images. But it is still a challenging task to detect
Neovascularization due to its abnormal random growth and
small size pattern. Recently, Deep learning techniques are
getting immensely popular due to their advancement in
Artificial Intelligence in Biomedical ImageProcessing.These
methods can be employed to traina CNN model todetectand
classify Neovascularization in fundus Images.
1) By the literature survey of the research papers, the
existing system methods which include image processing
and machine learning algorithms provide preferable
accuracy for detecting Neovascularization,butthisislimited.
The abnormal random growthpatternofNeovascularization

must be detected and identified so that the condition of
Neovascularization can be classified accurately.
2) The existing image processing techniques are better
suited for detecting the Neovascularization based on the
training data samples. The growth pattern still remains
challenging to detect and also it changes from patient to
patient. The growth pattern is abnormal and small. The
detection mechanism must be able to learn the behavior of
the growth pattern of blood vessels in order to detect and
classify it properly.
3) The existing systems with image processing techniques
can be improved by employing deep learning methods for
feature extraction and classification. The deep learning
models can be trained to learn the behavior of the random
growth pattern of the blood vessels formation and can be
able to detect and classify the condition of
Neovascularization by training the model with an ample
amount of training dataset and validating the model with a
validation dataset consisting of different growth patterns of
blood vessels. These models can be used to act as a base to
build the system for detecting and classifying the
Neovascularization from the fundus images.
M. C. S. Tang, S. S. Teoh, H. Ibrahim, and Z. Embong, in "A
Deep Learning Approach for the Detection of
Neovascularization in Fundus Images Using Transfer
Learning"[1] proposed a deep learning neural network that
is capable of detecting the Neovascularization in fundus
images. The performance of the transfer learning technique
is evaluated using four pre-trained networks AlexNet,
GoogLeNet, ResNet18, and ResNet50, and used to build the
proposed system for detecting the presence of
Neovascularization in fundus images. The combination of
ResNet18 and GoogLeNet pre-trained models is used in
building the proposed system for detecting the presence of
Neovascularization in fundus images.
S. C. Munasingha, K. K. Priyankara, R. G. Upasena, and A.
Ikeda in "A Novel Method of Detecting Neovascularization
Regions in Digital Fundus Photographs" proposed novel
methods to detect Neovascularization using image
processing techniques. The proposed system incorporates
image processing techniques and machinelearningmethods
on the fundus images to segment the regions with
Neovascularization. This segmentation can help find the
region of abnormal growth of blood vessels and can be able
to detect Neovascularization.
H. Huang, X. Wang, and H. Ma in "An Efficient Deep Learning
Network for Automatic Detection of Neovascularization in
Color Fundus Images" proposed a deep learning network to
automatically detect Neovascularization in fundus images.
The model is developed using FeaturePyramidNetwork and
Vovnet and the model is evaluated with fundus images from
the real-time cases. The experimental results with testing
showed that it has less training and test time and high
accuracy when compared with Mask R-CNN.
2. METHODOLOGY
The proposed network to detect and classify
Neovascularization in fundus images is developed using
deep learning methods and transfer learning approaches.
The popular pre-trained networks are evaluated on the
training and validation sets of fundus images and the model
with the best accuracy metrics is considered as the base
model for developing the proposed network for the
detection and classification. The proposed system employs
two tasks - Detection and Classification.
DETECTION:
The detection stage includes training the neural network
model to precisely detect Neovascularization in fundus
images. The abnormal random growth of blood vessels
makes it challenging to detect. The model should be capable
of learning the behavior of the random growth of the blood
vessels and should be able to detect the presence of
Neovascularization. The image processing techniques and
feature extraction methodologies can be combined with
neural networks to accurately detect the presence of
Neovascularization in fundus images. This detection stage
helps in further classifying the stage and condition of
neovascularization detected in the eyes of the patient.
CLASSIFICATION:
The classification stage of the proposed system enables the
model to predict the condition of Neovascularization in
fundus images. Neovascularization can lead to permanent
vision loss if not treated early. The early detection of
Neovascularization can help patients with their treatment
and recovery. There are 5 stages in Diabetic Retinopathy -
(Healthy/Normal, Mild, Moderate,Severe,andProliferative).
the presence of Neovascularization is detected in the
Detection stage and after the detection stage thecondition is
classified based on the growth of the abnormal bloodvessels
on the retina in the Classification stage. This helps the
patients to get the conventional and righttreatmentfortheir
condition detected through their fundus images. The model
should be able to learn the random growth patterns of the 5
stages and accurately detect the right condition of the
detected Neovascularization in the retina.
Fig -1: Stages of Neovascularization

SYSTEM ARCHITECTURE:
System architecture refers to the high-level design or
blueprint of the entire conceptual model of the proposed
system. It provides a structural overview of how the system
is built and organized. The proposed systemisbuiltontop of
the pre-trained neural networks using transfer learning
approaches. The underlying architecture for implementing
the proposed system is given below.
Fig -2: System Architecture
WORKING PRINCIPLE:
The fundus image dataset is collected according to the 5
stages - (Healthy, Mild, Moderate, Proliferate, andSevere). In
the image processing techniques, the ground truths of the
fundus images are obtained for the model training. The
dataset is split into training and validation sets to evaluate
the model performance over its training time. the proposed
network is developed using a transfer learning approach.
The pre-trained networks are loaded and trained over the
train and validation datasets and the model with the best
performance metrics is employed for developing the
proposed system to detectandclassifyNeovascularizationin
fundus images. The popular pre-trained networks used
include - Inception ResNetV2, DenseNet, ResNet50,
RestNet18, AlexNet, and VGG19 networks.Themodel isthen
loaded into a flask object to provide an interface for the
patient/user to interact with the system and upload their
fundus image to check their condition.
2.1 MODEL TRAINING:
I. IMAGE PROCESSING:
The fundus image dataset is collected according to the 5
stages - (Healthy, Mild, Moderate, Proliferate,andSevere). In
the image processing techniques, the ground truths of the
fundus images are obtained for the model training. The
dataset is split into training and validation sets to evaluate
the model performance over its training time. the proposed
network is developed using a transfer learning approach.
The pre-trained networks are loaded and trained over the
train and validation datasets and the model with the best
performance metrics is employed for developing the
proposed system to detectandclassifyNeovascularizationin
fundus images. The popular pre-trained networks used
include - Inception ResNetV2, DenseNet, ResNet50,
RestNet18, AlexNet, and VGG19 networks.Themodel isthen
loaded into a flask object to provide an interface for the
patient/user to interact with the system and upload their
fundus image to check their condition.
Fig -3: Abnormal blood vessels growth in the retina
Ground truth images are essential for evaluating the
performance of neovascularization detection algorithms or
systems. They are used as a reference for comparing the
results obtained from automated image analysis methods,
such as segmentation and classification algorithms, to
determine the accuracy, sensitivity, specificity, and other
performance metrics of the algorithms.
II. FEATURE EXTRACTION & CLASSIFICATION:
The feature extractionmethodcomprisesextractingrelevant
features or attributes from the segmented blood vessels.
This informationcanbeused todiscriminateandunderstand
normal blood vessels and their growth with the abnormal
growth of blood vessels in the retina. This can identify the
severity of the Neovascularization detected in the eye of the

patient. The features extracted could include vessel
diameter, vessel tortuosity, vessel branching patterns, and
other shape, intensity, and texture features.
Classification techniques are applied in the neural network
models to train in categorizing the segmented blood vessels
into 5 classes of diabetic retinopathy stages. The
Convolutional Neural network is trained on the image
datasets for classification.Theseneural networksaretrained
on labeled datasets to learnthepattern,growth,andfeatures
of the healthy and abnormal blood vessels and make
predictions on the new fundus images.
III. TRANSFER LEARNING:
Transfer learning is a machine learning technique that
leverages knowledge learned from one task or domain to
improve the performance of another, related task, or
domain. It involves training a pre-trained model, typically
trained on a large dataset for a different task, on a smaller
dataset, or on a different task, to benefit from the learned
representations and features. Transfer learning has gained
significant attention and popularity in recentyearsdueto its
ability to overcome data limitations and improve the
performance and efficiency of machine learning models in
various domains, including image recognition, natural
language processing, and speech recognition, among others.
One of the key advantages of transfer learning is that it
allows models to leverage the knowledge and
representations learned from a largedatasettoperform well
on a smaller dataset with limited labeled data. This is
especially beneficial in scenarios where obtaining large,
labeled datasets for training is challenging, time-consuming,
or expensive. By utilizing a pre-trained model, transfer
learning can help overcome the limitations of limited data,
resulting in more accurate and robust models. Transfer
learning can be implemented in different ways, depending
on the specific task and domain.
Fig -4: Transfer Learning Approach
The proposed system is constructed using the transfer
learning approach. In this approach,thepre-trainedmodel is
used as a fixed feature extractor, where the layersofthe pre-
trained model are used to extract relevant features from the
input data. These features are then fed into a separate
classifier or model for the target task. This allows the model
to benefit from the learned representations and features
from the pre-trained model,whichmayhavealreadylearned
generic patterns or high-level features that are transferable
to the target task. The deep architecture of layers of the pre-
trained neural network models helps the systemtolearn the
features from the datasets and understand the behavior of
the growth pattern of the blood vessels. However, it is
important to carefully select the appropriate pre-trained
model, consider the domain and task similarity between the
source and target tasks, and evaluate the performanceof the
transferred model to ensure its effectiveness for the specific
target task.
2.2 PRE-TRAINED NEURAL NETWORKS:
I. INCEPTION RESNETV2:
Inception ResNetv2 is a deep convolutional neural network
(CNN) architecture that combines the Inception moduleand
the residual connectionsfromResNet. InceptionResNetv2is
a pre-trained network that has been trained on a large
dataset to learn the features from images, making it a
powerful tool for a wide range of computer vision tasks. The
Inception module in Inception ResNetv2 is designed to
extract features at multiple scales by using convolutional
layers with different filter sizes (1x1, 3x3, and 5x5) and
pooling operations, and then concatenatingtheoutputs. This
allows the network to capture both local and global
contextual information, making it more robust to variations
in object scales and orientations. Residual connectionsallow
for the efficient training of very deep networksbyalleviating
the vanishing gradient problem, which can occur in deep
neural networks. Residual connections also help to improve
the accuracy and stability of the network during training by
allowing the network to learn both the residual and the
identity mapping, which can facilitate the flow of gradients
and information across the layers.
II. DENSENET:
DenseNet, short for Densely Connected Convolutional
Networks, is a deep Convolutional Neural Network (CNN)
architecture. It is known for its unique dense connectivity
pattern, where each layer receives input from all previous
layers, resulting in highly connected and densely packed
feature maps. DenseNet is a pre-trained network,meaningit
has been trained on a large dataset to learn meaningful
features from images, making it a powerful tool for various
computer vision tasks. The dense connectivity pattern in
DenseNet allows for efficient featurereuseandgradientflow

throughout the network, as each layer has access to the
feature maps of all previous layers. This dense connectivity
leads to better feature propagation, reduces the number of
parameters needed, and enhances the gradient flow during
training.
III. RESNET50:
ResNet50 is a popular pre-trained Convolutional Neural
Network (CNN) architecture. ResNet stands for Residual
Network, and 50 in ResNet50 refers to the depth of the
network, which consists of 50 layers. ResNet50 is knownfor
its residual connections, which allow for the efficient
training of deep networks and have been shown to improve
accuracy and stabilityduring training.ResNet50isorganized
into blocks, where each block contains multiple
convolutional layers followed by batch normalization and
activation functions, with residual connections across the
blocks. The basic building block of ResNet50 is the
bottleneck block, which consists of three convolutional
layers with different filter sizes (1x1,3x3,and1x1)toreduce
computational complexity, followed by batch normalization
and activation functions.
IV. RESNET18:
ResNet stands for Residual Network, and 18 in ResNet18
refers to the depth of the network, which consists of 18
layers. ResNet18 is known for its residual connections,
which allow for the efficient training of deep networks and
have been shown to improve accuracy and stability during
training. The basic building block of ResNet18 is the basic
block, which consists of two convolutional layers with the
same filter size (3x3), followed by batch normalization and
activation functions. The residual connections in ResNet18
skip one or more convolutional layers and directly connect
the input of the block to the output, allowing for efficient
feature propagation and reuse.
V. ALEXNET:
AlexNet is a pioneering pre-trained convolutional neural
network (CNN) architecture. AlexNet is known for its deep
architecture, use of rectifiedlinearunits(ReLU)asactivation
functions, and introduction of dropout regularization to
mitigate overfitting. AlexNet consists of five convolutional
layers followed by three fully connected layers. The
convolutional layers are responsible for extracting
hierarchical features from the input image, while the fully
connected layers are used for classification.Thearchitecture
of AlexNet is characterized by its depth, with a total of eight
layers, which was considered deep at the time of its
introduction. ReLU introduces non-linearity into the
network, allowing it to learn complex patterns and
representations from the input data.
VI. VGG19:
VGG19 is known for its simplicity and uniformity in design,
consisting of 19 layers, with 16 convolutional layers and 3
fully connected layers. VGG19 is widely used for image
recognition tasks, such as image classification and object
detection, due to its strong performance and ease of
implementation. The maincharacteristicofVGG19isits deep
architecture, with multiple convolutional layers stacked on
top of each other, followed by fully connected layers for
classification. VGG19usessmall convolutional filters(3x3)in
all its convolutional layers, which allows for a deeper
network with a smaller number of parameters compared to
larger filters. This leads to a more efficient and
computationally feasible architecture. VGG19 also has a
uniform design, where all the convolutional layers have the
same number of filters (64, 128, 256, or 512) and the same
padding (same) to maintain the spatial dimensions of the
input. This uniformity indesignmakesiteasiertoimplement
and fine-tune the network.
3. CHALLENGES
The proposed system to detect and classify
Neovascularization in fundus images is developed based on
deep learning methods. Various challenges are faced when
training a deep neural network on a fundus image dataset.
Some of those challenges include:
I. DATA AVAILABILITY:
Fundus images of patients with neovascularization are
relatively rare, making it challenging to obtain a large
dataset for training deep learning models. This can result in
models with limited generalization capabilities and reduced
accuracy and performance.
II. VARIABILITY IN GROWTH PATTERNS:
Neovascularization can manifest in different patterns, such
as fine or coarse vessels, which can vary in size, shape, and
appearance. This variability makes it challenging to develop
a robust deep learning model that can accurately detect
neovascularization across different patterns andstages. The
abnormal growth patterns of the blood vessels in the retina
makes it challenging to develop the model to learn the
behavior of the growth pattern and predict the condition.
III. NOISE AND ARTIFACTS:
Fundus images can be affected by various artifacts, such as
uneven illumination, blur, and noise, which can affect the
accuracy of neovascularization detection. Deep learning
models need to be robust to these artifacts to achieve
reliable detection performance.

IV. INTERPRETABILITY:
Deep learning models are often considered as "black box"
models, making it difficult to interpret and explain the
rationale behind their predictions. In the context of
Neovascularization detection, interpretability is important
for gaining trust and acceptance from clinicians and
stakeholders.
V. MODEL OVERFITTING:
Deep learning models are prone to overfitting, where they
may learn to memorize the training data instead of
generalizing from it. This can result in poor model
performance on unseen data, including fundus images with
Neovascularization.
VI. COMPUTATIONAL RESOURCES:
Deep learning models typically require significant
computational resources, including powerful hardware and
large amounts of memory, for training and inference.Access
to such resources may be limited in some settings, posing
challenges in developing and deploying deep learning
models for neovascularization detection in resource-
constrained environments.
4. TESTING AND RESULTS
The pre-trained networks are trained on the fundus image
dataset. The image dataset is divided into training and
validation dataset and each pre-trained network is trained
on those datasets. The validation dataset is used to evaluate
the performance metrics and accuracy of the model in
detecting and classifying the Neovascularization in fundus
images. The model with the best accuracymetricsisadopted
for constructing the deep neural network for detecting and
classifying the stage of Neovascularization.
The pre trained networks employed for developing the
proposed network are trainedover3000+imagesoftraining
data and validated over 300+ images of validation dataset.
The network which performs best under optimal conditions
and gives a better accuracy for less training time is
considered for building the proposed system. The testing
stage here includes the accuracy testing of each pre trained
network to find the best model architecture to build the
system.
4.1 MODEL COMPARISION:
PRE-TRAINED MODEL MODEL ACCURACY
Inception ResNetv2 86%
DenseNet 74%
ResNet50 72%
ResNet18 71%
AlexNet 72%
VGG19 72%
Table-1: Performance of Neural Networks
The Inception ResNetV2 Model has shown the highaccuracy
of 86% with less training time over a less training dataset
compared to other models. This pre trained network
Inception ResNetV2 isusedforbuildingtheproposedsystem
for detecting and classifying the Neovascularization.
4.2 RESULTS
From the model training and validation, the Inception
ResNetV2 model has better accuracy of prediction when
compared to other CNN Models. This forms the base for
developing the proposed system and the proposed model is
loaded into a Python Flask object to create a web interface
for the users to upload fundus images and classify its stage.
Fig -5: Prediction Result on New Image
5. CONCLUSION
The Neovascularization detection and classification of
diabetic retinopathy in fundus images using deep learning
methods has a great impact in advancing the field of
ophthalmology. Through leveraging pre-trainedmodelsand
utilizing deeplearningtechniques, researchersandclinicians
have been able to achieve remarkable results in automating
the detection of Neovascularization, a critical feature in
various retinal diseases such as diabetic retinopathy.
Transfer learning has been found to beaneffectiveapproach
in overcoming the limitations of limited annotated data and
reducing the need for large datasets, which are often
challenging to obtain in medical imaging. By leveraging pre-
trained models, such as convolutional neural networks
(CNNs), that are trained on large datasets from other tasks,
transfer learning allows for the extraction of meaningful
features from fundusimages andenablesthedevelopmentof
accurate and robust neovascularization detection models
with relatively smaller datasets.

The integration of transfer learning with deep learning
approaches has further improved the accuracy and
robustness of neovascularization detection models. By fine-
tuning pre-trained CNNs with domain-specificdata,transfer
learning allows for the adaptation of the model to the
specific characteristics of the fundus images, leading to
improved performance in detecting Neovascularization.
The performance of six pre-trained convolutional neural
networks, which are Inception ResNetV2, DenseNet,
ResNet50, ResNet18, AlexNet, and VGG19, was investigated
for building the best CNN model for developing the
neovascularization detection model through transfer
learning. Evaluation results show that the transfer learning
approach yields superior performance. The Inception
ResNetV2 Model architecture is considered the best model
for developing the proposed systemandthemodel istrained
and loaded into a flask object to classify the fundus images
into 5 classes of Neovascularization conditions (Healthy,
Mild, Moderate, Severe, and Proliferate).
6. FUTURE SCOPE
 More fundus images can be added to the training and
validation datasets to increase the accuracyofdetection
and classification. The fundus images of patientsin real-
time can be served as new validation sets to fit the
model training and improve its accuracy of prediction.
 The detection model incorporated on the website gives
the patients the chance to check their condition and get
immediate treatment.
 The model training over large fundus image datasets to
build a much more accurate detection model could be
efficient when the system is connected to a GPU,
allowing it to run smoothly and reduce the model
training time.
 The detection model can also be embedded into an IOT
device for real-time detection and classification
purposes in the Medical Industry.
 Continued research and development in this field,along
with careful validation and standardization, will
contribute to the further advancement and integration
of these approaches into clinical practice, ultimately
benefiting patients and improving the management of
retinal diseases.
REFERENCES
[1] M. C. S. Tang, S. S. Teoh, H. Ibrahim, and Z. Embong, "A
Deep Learning Approach for the Detection of
Neovascularization in Fundus Images Using Transfer
Learning," in IEEE Access, vol. 10, pp. 20247-20258,
2022, doi: 10.1109/ACCESS.2022.3151644.
[2] S. C. Munasingha, K. K. Priyankara, R. G. Upasena, and A.
Ikeda, "A Novel Method of Detecting Neovascularization
Regions in Digital Fundus Photographs," 2022 IEEE 4th
Eurasia Conference on Biomedical Engineering,
Healthcare and Sustainability (ECBIOS), 2022, pp. 1-4,
doi: 10.1109/ECBIOS54627.2022.9945014.
[3] H. Huang, X. Wang, and H. Ma, "An Efficient Deep
Learning Network for Automatic Detection of
Neovascularization in Color Fundus Images,"202143rd
Annual International Conference of the IEEE
Engineering in Medicine & Biology Society (EMBC),
2021, pp. 3688-3692, doi:
10.1109/EMBC46164.2021.9629572.
[4] M. Z. Khan and Y. Lee, "Retinal Image Analysis to Detect
Neovascularization usingDeepSegmentation,"20214th
International Conference on Information andComputer
Technologies (ICICT), 2021, pp. 110-114, doi:
10.1109/ICICT52872.2021.00026.
[5] K. Firdausy, O. Wahyunggoro, H. A. Nugroho and M. B.
Sasongko, "A Study on Recent Developments for
Detection of Neovascularization," 2019 11th
International Conference on Information Technology
and Electrical Engineering (ICITEE), 2019, pp. 1-6, doi:
10.1109/ICITEED.2019.8929941.
[6] H. Huang, H. Ma, and W. Qian, "Automatic Parallel
Detection of Neovascularization from Retinal Images
Using Ensemble of Extreme Learning Machine," 2019
41st Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC),
2019, pp. 4712-4716, doi:
10.1109/EMBC.2019.8856403.

A Deep Learning Approach for the Detection and Identification of Neovascularization in Fundus Images

More Related Content

Similar to A Deep Learning Approach for the Detection and Identification of Neovascularization in Fundus Images (20)

More from IRJET Journal (20)

Recently uploaded (20)

A Deep Learning Approach for the Detection and Identification of Neovascularization in Fundus Images