Journal of Computer Science

Original Research Paper

Diabetic Retinopathy Detection using Deep Learning

Techniques
Aswin Shriram Thiagarajan, Jithendran Adikesavan,
Santhi Balachandran and Brindha Ganapathyagraharam Ramamoorthy

SASTRA Deemed University, Thanjavur, India

Article history Abstract: Diabetic Retinopathy is a type of eye condition induced by

Received: 31-07-2019 diabetes, which damages the blood vessels in the retinal region and the area
Revised: 08-10-2019 covered with lesions of varying magnitude determines the severity of the
Accepted: 13-03-2020 disease. It is one of the most leading causes of blindness amongst the
Corresponding Author:
employed community. A variety of factors are observed to play a role in a
Santhi Balachandran person to get this disease. Stress and prolonged diabetes are two of the most
SASTRA Deemed University, critical factors to top the list. This disease, if not predicted early, can lead to
Thanjavur, India a permanent impairment of vision. If predicted in advance, the rate of
Email: [email protected] impairment can be brought down or averted. However, it is not easy to
detect the presence of this disease, given the time-consuming and tedious
process of diagnosis. Presently, digital color photographs are evaluated
manually by trained clinicians to observe the presence of lesions caused due
to vascular abnormalities, which is the major effect of Diabetic
Retinopathy. This method, although it is pretty accurate, proves to be
costly. The delay brings out the need to automate the diagnosing, which
will, in turn, have a significant positive impact on the health sector. In
recent times, the adoption of AI in disease diagnosis has ensured promising
and reliable results and this serves as the motivation for this journal. The
paper employs Deep learning methodologies for automatic detection of
Diabetic Retinopathy, resulting in a maximum accuracy of 80%, as
compared to traditional Machine learning approaches giving only a
maximum accuracy of 48% on the same IRDiR Disease Grading Dataset
(413 images with 5 levels of DR-Training set; 103 images with 5 levels of
DR-Test set). The data set contains digital fundus images of different levels
of Diabetic Retinopathy in discrete frequency distributions.

Keywords: Diabetic Retinopathy, CNN, Deep Learning, Feature

Engineering, Artificial Intelligence

Introduction diagnoses the problem in no time with the past history and
the data tagged with it. Diabetic Retinopathy is one such
Computer-aided medicine, digital health consultancies, problem that the article tries to address.
health monitoring systems are some of the recent Diabetic Retinopathy is caused as an outcome of
buzzwords that are being viral nowadays. Common people diabetes damaging the blood vessels present on the
are given the luxury of treatment and diagnosis from their interior of the retina, leading to leakage of fluids and
homes with a single tap, thanks to the connectivity and blood into the tissues neighboring it. The leakage
computational infrastructure that has given rise to the develops soft exudates (also referred to as cotton wool
digital era we live in. While common illness or mild spots), hard exudates, microaneurysms and hemorrhages.
infections are very curable without consulting a qualified It is the primary source of vision impairment amongst
doctor, some severe infections still require lots of effort the working population.
from the medical community, wherein technology can This paper aims at developing a deep learning
only aid the process but cannot act independently. based automated DR detection model from a frugal
With the development of Artificial intelligence, amount of data. The performance of DL models will
technology independently assesses the patient’s health and be juxtaposed with other traditional algorithms

© 2020 Aswin Shriram Thiagarajan, Jithendran Adikesavan, Santhi Balachandran and Brindha Ganapathyagraharam
Ramamoorthy. This open access article is distributed under a Creative Commons Attribution (CC-BY) 3.0 license.
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

reinforced with and without various image processing proven to provide superior performance much better than
techniques (like filtering, smoothing, etc.,). DL being traditional ML algorithms. Xu et al. (2017) used deep
a black-box approach is augmented by inputting CNN with data augmentation techniques to classify
feature engineered images to provide a head start in fundus images as DR and normal, acquiring accuracy
the performance. of approximately 95%. This approach demonstrated to
be more efficient than the XGBoost model with feature
Literature Survey extracted inputs. Islam et al. (2018) and Pratt et al.
(2016). had also developed CNN based models to
Earlier researches on Diabetic Retinopathy detection
identify Microaneurysms (MA) using Kaggle datasets,
used SVM to classify the Non-Proliferative Diabetic
achieving a sensitivity of 95% and above. Wan et al.
Retinopathy (NPDR) into different grades. Carrera et al.
(2018) incorporate transfer learning methodology on
(2017) the implementation began with the extraction of
popular CNN architectures for enhanced performance in
features like Blood vessel density, Number of
DR image classification.
Microaneurysms and density of hard exudates. The
Similar to the above approach, (Maya and Adarsh,
classification was in two phases, namely, NPDR detection
2019) use CNN to detect bright lesions in the fundus
and grade classification in the NPDR. The results were
images. The main difference between (Andonová et al.,
benchmarked against the Decision Trees (DT) approach, 2017). method and the former is that, the image is
obtaining an average accuracy of 85% and a maximum converted into smaller blocks instead of using the
sensitivity of 95%. Another paper on proliferative DR whole image, followed by preprocessing and
(Adarsh and Jeyakumari, 2013). leverages Multiclass transformation (like Adaptive Histogram Equalization,
SVM (one against all) to classify the severity of DR into Gaussian Noise, Gray scaling, green channel) and
normal, mild, moderate, severe and proliferative. The subsequently fed into the CNN architecture with 4
approach included manual feature extraction, which was convolution and 2 fully connected layers for
extensively based on morphological operations and classification. Validation is done by comparing against
features for the model were selected after performing a expert marked fundus images.
statistical significance test. The average accuracy and Hemanth et al. (2019) use hybrid techniques using
sensitivity were 95% and 91%, respectively. Deep Learning and traditional image processing
Rubya and Pintu (2019). incorporated a fuzzy logic methods for DR diagnosis from retinal fundus images.
classifier to detect different stages of DR, resulting in up The research findings above prove Deep Learning
to 95% accuracy. Jothi and Jayaram (2019) developed a techniques supersede any other approaches when it
method to classify the blood vessels that potentially lead comes to automated and efficient detection of DR and
to DR and other retinal vascular defects using the (Thanati et al., 2019). consolidates some of the recent
Firangi filter technique. Both the above papers had approaches. The review paper (Bellemo et al., 2019)
used retinal fundus image data from DIARETDB and extensively discuss the increased usage of emerging
STARE databases. technologies like Artificial Intelligence in the national
Karim et al. (2019) used MATLAB’s Neural screening of Diabetic Retinopathy around the globe.
Network Pattern recognition tool to investigate DR The paper presented here also deals with CNN based
symptoms by detection of micro-aneurysm. The analysis approach incorporating tuning of hyperparameters,
was benchmarked against the traditional ML methods whereas emphasizing primarily on having fewer layers
such as Naïve Bayes and SVM. and lesser trainable parameters to achieve better and
The recent progress involves research on reliable results. Although the papers referenced above
automated computer-aided diagnosis using Deep had employed deep learning paradigms for detecting
Learning approaches to detect and grade Diabetic DR, they either consisted of deeper layers with great
Retinopathy in advance, from the digital fundus number of parameters or had consumed huge quantities
images. Balyen and Peto (2019) discuss the general of datasets. The procedure will be discussed in detail in
influence of AI, ML and DL on the field of the subsequent sections.
ophthalmology. Cheung et al. (2019). suggest the
potential of DL in the detection of DR and Diabetic Methodology
Macular Edema (DME) from optical coherence
tomography and digital fundus images. Recent trends This paper mainly concentrates on the application of
in research show that DR detection is done prevalently Convolution Neural Networks in the diagnosis of
using Deep Learning Techniques. The reason being, Diabetic Retinopathy. The salient features of the CNN or
ease of obtaining results with less preprocessing any Deep Learning paradigm are automatic feature
efforts. Papers employing the CNN approach has extraction and efficient computation. These

306
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

advantages of CNN provided motivation to pursue this The input images, as seen from the algorithm’s
approach in this paper. perspective, are just matrices of pixel values of
A CNN model is composed of a convolution layer, a dimension (height, width, number of channels). Here,
pooling layer and an activation layer in multiple channels refer to RGB layers.
combinations of stacks. The input data fed into the First, pre-processing is the reshaping of input data.
network will be convolved with various filters. These The images acquired are of size 4200×2800 pixels and
filters are akin to traditional image processing filters, occupy 0.5 mb of memory space. This demands too
except that they are learned automatically instead of much of RAM usage, resulting in slower computation.
being defined explicitly. Hence, are downsized to 256×256 dimension.
The 413 images from the training set were used in the Images from the training set will have inherent
development and validation of the model. The model variance and direct feeding of input data to the CNN will
was tested on the 103 images test set. result in slower convergence of solution and unexpected
The problem architecture is split into multiple phases, results. Hence, following the reshaping, the image needs
as shown in Fig. 1. to be normalized.
Finally, the processed images are split into train and
Pre-Processing validation set of 4:1 ratio (80%: 20%). This validation
Images from the training dataset (samples in Fig. 2) set is utilized to gauge the capacity of the model.
are read using the OpenCV library. This library is Data Augmentation
essential for image processing, as it contains rich built-in
features for rapid processing. Any deep learning model is data thirsty and in order
to expand data, a data augmenter is used to generate
more data without the need to collect more actual data.
Input image
Keras library has inbuilt functionality to perform this
operation. Some augmentation techniques include
image rotation (to induce rotational invariance),
horizontal flipping, scaling, zooming in, cropping and
translation. A very recent paper (Frid-Adar et al., 2018)
Image pre-processing uses state of the art Generative Adversarial Networks
(GANs) to create new data. The next section discusses
the methodologies in designing a Deep Convolution
Neural Network.

Image augmentation

Deep CNN design

Until best results

Hyperparameters selection

Model performance

Fig. 1: Proposed architecture flow diagram Fig. 2: iRDRiD data set samples

307
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

This paper involves some important data same dimensional matrix. In a Deep CNN context, the
augmentation procedures, namely, image rotation with a filters/kernels are parameters that are learned by a
range of 25° either side, image shearing with a range of suitable backpropagation algorithm.
0.2 radians, zooming of image with a range of 0.2, The pooling layer subsequently decreases the
horizontal flipping, random shifting of images in the dimension of the data, as represented in Fig. 4.
height axis with a range of 0.1, random shifting of Conventionally max pooling and average pooling are
images in the width axis with a range of 0.1, while used in CNN to obtain more stable results.
shifting fill mode of ‘nearest’ was chosen and image Batch Normalization is the layer preceding activation
normalization across features. layer to reduce the internal covariant shift (Ioffe and
Deep CNN Design Szegedy, 2015). This aids in increased speed of training.
The activation layer performs the role of non-linear
Following the pre-processing step, the design of function mapping that enables the learning of complex
CNN is carried out. As hinted before, it is a combination features. It only fires those neurons that observe a
of stacks of layers to perform the necessary particular feature in that layer (Feature maps in every
mathematical computation Fig. 6. layer Fig. 6. Rectified Linear Unit (ReLU), Sigmoid,
The convolution layer is the primary entity, that Scaled Exponential Linear Units (SeLU) are some of the
enables an image to undergo transformation to learn its activation functions used in this paper. Softmax is
features. The convolution operation is performed, as generally put to use as the last layer that outputs a multi-
shown in Fig. 3. This operation results in a lower or class predicted label.

0 1 1 1 0 0 0

0 0 1 1 1 0 0 1 4 3 4 1

0 0 0 1 1 1 0 1 0 1 1 2 4 3 3

0 0 0 1 1 0 0 * 0 1 0 1 2 3 4 1

0 0 1 1 0 0 0 1 0 1 1 3 3 1 1

0 1 1 0 0 0 0 3 3 1 1 0

1 1 0 0 0 0 0

Fig. 3: Convolution Method; I-Image, K-Kernel.

Single depth slice

x 1 1 2 4

5 6 7 8 Max pool with 2×2 6 8

filters and stride 2

3 2 1 0 3 4

1 2 3 4

Fig. 4: Maxpool with stride 2

308
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

Dropouts are regularization techniques, used in Gray-Level Co-Occurrence Matrix (GLCM) features,
diminishing the overfitting of training data. A dropout divided into k equal subsets, was used in the model
layer is introduced in the architecture, which randomly training. That is, each ML model is trained k times with k-
prunes off neurons during training. 1 buckets as the training set and the remaining 1 bucket as
Several of the above building blocks constitute a the validation set. The output evaluation metric, accuracy,
CNN model for image classification. The output from is considered for analysis. Two different sets of
the stack of CNN layers will be flattened, converted from approaches were implemented and it is as follows.
(height, width, number of channels) vector into a vector of GLCM feature extraction from the raw input image
shape (height * width * number of channels, 1). A fully without performing any processing (application of any
connected layer follows the previous flattened layer. filter, particularly Gaussian filter). Box plots were
Once the layers are constructed, the weights plotted to explore the accuracy ranges of the model.
(parameters) and bias of the neurons are randomly From Fig. 5a, it can be inferred that a simple Logistic
initialized and the model is trained to learn them. A Regression model performed better than the rest of the
relevant loss function like Binary Cross Entropy was models, with Random Forest performing equally well.
defined to measure the error in prediction. Optimizers Their average accuracy from Table 1, will
enable the learning of parameters that minimizes the substantiate the above observation, that the LR model
cost and decides the speed of learning. Effects of
has the maximum average accuracy of approximately
optimizers Adam, Adagrad and RMS Prop were
48%. The standard deviation denotes the
applied in this paper and their analysis results are
generalization capability of a model. Lower the value,
discussed later in this journal.
better generalization is achieved.
Hyperparameter Tuning and Selection b. GLCM feature extraction from the input image
after applying the Gaussian filter. Box plots were plotted
From the previous component, modifying certain to explore the accuracy ranges of the model. From Fig.
entities like the number of CNN blocks, activation 5b, it can be inferred that a simple Logistic Regression
functions used, filter size, type of pooling, types of model performed better than the rest of the models, with
optimizers and their initial learning rates, resulted in Random Forest performing equally well.
varying performance of the model Table 3 validated Their average accuracy from Table 2, will justify
against the validation/test set. The hyperparameter the above observation, that the LR model has the
settings that gave the best results were considered for maximum average accuracy of approximately 49%.
the final model Table 4. Referring to other journals The standard deviation denotes the generalization
and analysis of the dataset helped in funneling down capability of a model. Lower the value, better
the possible combinations of hyperparameters that can generalization is achieved.
be experimented.
Prediction and Validation Table 1: Average and standard deviation of cross validation
(10-fold) score of ML models without Gaussian
After the model was trained with the best possible filtering (GLCM features)
accuracy, it is ready to be used for the prediction of Algorithm Accuracy % Std deviation
unseen data. During the training phase, the datasets were Logistic Regression 48.236 0.099
fed into the model in mini-batches of 32 images every Linear Discriminant Analysis 33.400 0.083
iteration. The weights were learned for 20 epochs. The k-Nearest Neighbors 43.926 0.069
testing was performed on the test dataset and accuracy Decision Tree 31.280 0.079
Random Forest 47.176 0.083
was used as the performance evaluation metric. Detailed
Naive Bayes 21.266 0.057
analysis of deep learning model results and comparison Support Vector Machine 33.407 0.087
with traditional ML algorithms are discussed in the
subsequent section.
Table 2: Average and standard deviation of cross validation
(10-fold) score of ML models with Gaussian filtering
Results and Discussion (GLCM features)
Algorithm Accuracy % Std deviation
Initial attempts involved modeling the dataset using
Logistic Regression 48.784 0.084
standard Machine learning techniques like Logistic
Linear Discriminant Analysis 35.028 0.080
Regression, Linear Discriminant Analysis, k-NN, k-Nearest Neighbors 41.771 0.096
Decision Tree, Random Forest, Support Vector Decision Tree 37.475 0.087
Machines and Naïve Bayes. In this paper, an ensemble Random Forest 45.811 0.074
learning method is used in ML modeling with a k-fold Naive Bayes 19.132 0.067
cross-validation approach. The dataset with extracted Support Vector Machine 33.407 0.087

309
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

0.7

0.6

0.5
Accuracy

0.4

0.3

0.2

0.1
LR LDA KNN CART RF NB SVM
ML models
(a)

0.6

0.5

0.4
Accuracy

0.3

0.2

0.1

LR LDA KNN CART RF NB SVM

ML models
(b)

Fig. 5: (a) Cross validation (10-fold) score of ML models without gaussian filtering (GLCM features); (b) Cross validation (10 fold)
score of ML models with gaussian filtering (GLCM features)

The above results imply standard ML techniques did not Maximum accuracy of 80% was attained in the
fetch good accurate predictions and it did not improve architecture containing a total of 5 CNN Layers (4 with
despite executing feature engineering (application of filters) ReLU activation, 1 with Softmax activation), trained using
on the dataset. Since the ML models failed to deliver results Adam optimizer initialized with a learning rate of 10−3 for
for this dataset, Deep Learning techniques were faster convergence of results. The resulting architecture
implemented. Table 3 provides the outcomes of different contains fewer parameters compared to the others and the
experimented architectures. proposed configuration consisting of a unique combination

310
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

of optimizers, loss function and learning rate proves to be the best solution for this scarce dataset.

Output layer (softmax activation)

0
Input image

Fully connected layer

0

Flattened layer
Conv L1 Conv L2 Conv L3 Conv L4
ReLU ->BN ReLU ->BN ReLU ->BN ReLU ->BN 1
-> MP->D -> MP->D -> MP->D
0

Fig. 6: CNN model architecture diagram with feature maps at every CNN layer

Table 3: CNN model accuracy with different hyperparameter combinations

S.No No. of layers Layers information Optimizer Learning rate Loss function Accuracy %
1 6 ReLU, 4x(Sigmoid), Softmax Adam 0.0001 Categorical Cross entropy 36.145
2 6 5x(ReLU), Softmax RMSProp 0.001 Binary Cross entropy 69.675
3 6 ReLU, 4x(SeLU), Softmax Adam 0.001 Binary Cross entropy 71.084
4 8 7x(ReLU), Softmax Adam 0.001 Binary Cross entropy 75.422
5 6 5x(ReLU), Softmax Adagrad 0.001 Binary Cross entropy 77.108
6 5 4x(ReLU), Softmax Adam 0.001 Binary Cross entropy 80.036

Table 4: CNN model architecture summary

Layer (type) Output shape Number of parameters
conv2d_1 (Conv2D) (None, 256, 256, 32) 896
activation_1 (Activation) (None, 256, 256, 32) 0
batch_normalization_1 (Batch1) (None, 256, 256, 32) 128
max_pooling2d_1 (MaxPooling2) (None, 85, 85, 32) 0
dropout_1 (Dropout) (None, 85, 85, 32) 0
conv2d_2 (Conv2D) (None, 85, 85, 64) 18496
activation_2 (Activation) (None, 85, 85, 64) 0
batch_normalization_2 (Batch2) (None, 85, 85, 64) 256
max_pooling2d_2 (MaxPooling2) (None, 42, 42, 64) 0
dropout_2 (Dropout) (None, 42, 42, 64) 0
conv2d_3 (Conv2D) (None, 42, 42, 128) 73856
activation_3 (Activation) (None, 42, 42, 128) 0
batch_normalization_3 (Batch3) (None, 42, 42, 128) 512
conv2d_4 (Conv2D) (None, 42, 42, 128) 147584
activation_4 (Activation) (None, 42, 42, 128) 0
batch_normalization_4 (Batch4) (None, 42, 42, 128) 512
max_pooling2d_3 (MaxPooling3) (None, 21, 21, 128) 0
dropout_3 (Dropout) (None, 21, 21, 128) 0
flatten_1 (Flatten) (None, 56448) 0
dense_1 (Dense) (None, 1024) 57803776
activation_5 (Activation) (None, 1024) 0
batch_normalization_5 (Batch5) (None, 1024) 4096
dropout_4 (Dropout) (None, 1024) 0
dense_2 (Dense) (None, 5) 5125
activation_6 (Activation) (None, 5) 0
Total params: 58,055,237
Trainable params: 58,052,485
Non-trainable params: 2,752

311
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

The inherent feature extraction quality of a automatic hyperparameter tuning methods like meta-
convolution neural network has produced greater than learning, evolutionary search in parameter space.
70% accurate results in most of the cases with a simpler Additional reinforcements can be made with the
architecture and no additional feature engineering phase. confluence of highly performing Deep Learning and
The primary rationale behind this paper is the proposal Machine Learning models.
of a light-weight and straightforward Deep Learning
architecture and benchmark the performance against the Acknowledgement
conventional Machine Learning solutions. This model
can be deployed in low memory devices like basic This research was carried out at SASTRA University,
smartphones with less computational complexity and School of Computing, Research Labs and we would like to
achieving higher throughput. extend our gratitude to the institution for providing us with
the necessary computing resources and guidance.
Conclusion
Author’s Contributions
From the results obtained, it can be inferred that
deep learning models have been successful in Aswin Shriram Thiagarajan: Coding of CNN,
capturing the underlying pattern in the data. Hence, Content writing and formatting.
the increased performance of these models in Jithendran Adikesavan: Coding of CNN, Content
comparison to the machine learning models. writing and Dataset selection.
A variety of machine learning models were tried Brindha Ganapathyagraharam Ramamoorthy:
and the corresponding results were plotted. Random Structural design of paper.
Forest models capture the underlying distribution of Santhi Balachandran: Algorithm selection and
the images in the data set but are not able to match the content revision.
performance given by deep learning models. The
maximum accuracy obtained from a ML model was Ethics
48% after cross-validation and feature engineering,
whereas the average accuracy of DL models tried, This article is original and contains unpublished
were greater than 70%, with a maximum reaching material. The corresponding author confirms that all of
80% without much complexity. This reports a the other authors have read and approved the manuscript
significant 67% performance improvement with lesser and no ethical issues involved.
pre-processing efforts, leaving the deep learning
model a clear preference over other traditional References
algorithms for automated DR detection.
The originality in our approach lies in the frugal Adarsh, P. and D. Jeyakumari, 2013. Multiclass SVM-
designing of a lightweight Deep CNN architecture (less based automated diagnosis of diabetic
than 6 layers) and suitable hyperparameter settings with retinopathy. Proceedings of the International
efficient usage of the IRDiR dataset, which had very little Conference on Communication and Signal
data to be used in a data-intensive deep learning approach. Processing, Apr. 3-5, IEEE Xplore Press,
In comparison, other research papers used around 80000 Melmaruvathur, India, pp: 206-210.
images to develop a deep learning-based solution. The DOI: 10.1109/iccsp.2013.6577044
emphasis of our solution is a trade-off between data size Andonová, M., J. Pavlovičová, S. Kajan, M. Oravec
and performance using a lightweight Deep Learning model. and V. Kurilová, 2017. Diabetic retinopathy
Thus, automated diagnosing of Diabetic screening based on CNN. Proceedings of the
Retinopathy with the help of digital fundus images International Symposium ELMAR, Sept. 18-20,
experimented, with a noble cause of preventing the IEEE Xplore Press, Zadar, Croatia, pp: 51-54.
disease which is spread across the community. It will DOI: 10.23919/ELMAR.2017.8124433
assist doctors in early diagnosis and suggestion of Balyen, L. and T. Peto, 2019. Promising artificial
precautionary treatment for vision impairment. Further, intelligence-machine learning-deep learning
it inculcates confidence in patients on the curability of algorithms in ophthalmology. Asia-Pacific J.
this disease with a higher success rate. Ophthalmol., 8: 264-272.
DOI: 10.22608/APO.2018479
Future Scope Bellemo, V., G. Lim, T. Rim, G. Tan and C. Cheung et al.,
2019. Artificial intelligence screening for diabetic
Further, the solution can be improvised to achieve retinopathy: The real-world emerging application.
exceptional accuracy and minimize the categorization of Curr. Diab. Rep., 19: 72-72.
false positives and false negatives, with the employing DOI: 10.1007/s11892-019-1189-3

312
Aswin Shriram Thiagarajan et al., / Journal of Computer Science 2020, 16 (3): 305.313
DOI: 10.3844/jcssp.2020.305.313

Carrera, E.V., A. González and R. Carrera, 2017. Maya, K.V. and K.S. Adarsh, 2019. Detection of retinal
Automated detection of diabetic retinopathy using lesions based on deep learning for diabetic
SVM. Proceedings of the IEEE 24th International retinopathy. Proceedings of the 5th International
Conference on Electronics, Electrical Engineering Conference on Electrical Energy Systems, Feb.
and Computing, Aug. 15-18, IEEE Xplore Press, 21-22, IEEE Xplore Press, Chennai, India.
Cusco, Peru, pp: 1-4. DOI: 10.1109/ICEES.2019.8719242
DOI: 10.1109/INTERCON.2017.8079692 Pratt, H., F. Coenen, D. Broadbent, P.S Harding and Y.
Cheung, C., F. Tang, D. Ting, G.S.W. Tan and T.Y. Zheng, 2016. Convolutional neural networks for
Wong, 2019. Artificial intelligence in diabetic eye diabetic retinopathy. Proc. Comput. Sci., 90: 200-205.
disease screening. Asia-Pacific J. Ophthalmol., 67: DOI: 10.1016/j.procs.2016.07.014
1004-1009. DOI: 10.4103/ijo.IJO_1989_18 Rubya, A. and C.S. Pintu, 2019. Automatic lesions
Frid-Adar, M., E. Klang, M. Amitai, J. Goldberger and detection and classification of diabetic retinopathy
H. Greenspan, 2018. Synthetic data augmentation using fuzzy logic. Proceedings of the International
using GAN for improved liver lesion classification. Conference on Robotics, Electrical and Signal
Proceedings of the 15th International Symposium on Processing Techniques, Jan. 10-12, IEEE Xplore
Biomedical Imaging, Apr. 4-7, IEEE Xplore Press, Press, Dhaka, Bangladesh, Bangladesh, pp: 527-532.
Washington, DC, USA, pp: 289-293. DOI: 10.1109/ICREST.2019.8644123
DOI: 10.1109/ISBI.2018.8363576 Thanati, H., R.J Chalakkal and W.H. Abdulla, 2019. On
Hemanth, D.J., O. Deperlioglu and U. Kose, 2019. An deep learning based algorithms for detection of
Enhanced diabetic retinopathy detection and diabetic retinopathy. Proceedings of the
classification approach using deep convolutional International Conference on Electronics,
neural network. Neural Comput. Applic. Information and Communication, Jan. 22-25, IEEE
Ioffe, S. and C. Szegedy, 2015. Batch normalization: Xplore Press, Auckland, New Zealand.
Accelerating deep network training by reducing DOI: 10.23919/ELINFOCOM.2019.8706431
internal covariate shift. Proceedings of the 32nd Wan, S., Y. Liang and Y. Zhang, 2018. Deep
International Conference on International convolutional neural networks for diabetic
Conference on Machine Learning, Jul. 06-11, retinopathy detection by image classification.
JMLR.org, Lille, France, pp: 448-456. Comput. Electrical Eng., 72: 274-282.
Islam, S.M., M.M. Hasan and S. Abdullah, 2018. Deep DOI: 10.1016/j.compeleceng.2018.07.042
learning based early detection and grading of Xu, K., D. Feng and H. Mi, 2017. Deep convolutional
diabetic retinopathy using retinal fundus images. neural network-based early automated detection
Jothi, A. and S. Jayaram, 2019. Blood Vessel Detection in of diabetic retinopathy using fundus
Fundus Images Using Frangi Filter Technique. 1st image. Molecules, 22: 2054-2054.
Edn., Smart Innovations in Communication and DOI: 10.3390/molecules22122054
Computational Sciences. Springer, Singapore,
ISBN-13: 978-981-10-8970-1, pp: 49-57.
Karim, T., M.S. Riad and R. Kabir, 2019. Symptom
analysis of diabetic retinopathy by micro-Aneurysm
detection using NPRTOOL. Proceedings of the
International Conference on Robotics, Electrical and
Signal Processing Techniques, Jan. 10-12, IEEE
Xplore Press, Dhaka, Bangladesh, pp: 606-610.
DOI: 10.1109/ICREST.2019.8644439

313