Available online at www.sciencedirect.

com
ScienceDirect
ScienceDirect
Procedia Computer Science 00 (2022) 000–000
Available online at www.sciencedirect.com www.elsevier.com/locate/procedia
Procedia Computer Science 00 (2022) 000–000
www.elsevier.com/locate/procedia
ScienceDirect
Procedia Computer Science 218 (2023) 1055–1066

International Conference on Machine Learning and Data Engineering

International Conference on Machine Learning and Data Engineering
Diabetic Retinopathy Grading From Color Fundus Images:
Diabetic Retinopathy
An Autotuned Grading From Color
Deep Learning Fundus Images:
Approach
An Autotuned Deep Learning Approach
Athira T R , Jyothisha J Nair*
Athira T R , Jyothisha J Nair*
Department of Computer Science and Engineering, Amrita Vishwa Vidyapeetham, Amritapuri, India.

Department of Computer Science and Engineering, Amrita Vishwa Vidyapeetham, Amritapuri, India.
Abstract

Abstract
Diabetic Retinopathy (DR) is a consequence of long-term diabetes which affects the eyes. It causes affliction to the veins in the
eyes causing raptures on the retina which can impact vision. If the condition is failed to be detected at an early stage, it can lead to
Diabetic
completeRetinopathy (DR)conventional
vision loss. The is a consequence of long-term
diagnosing cycle ofdiabetes
DR using which affects
fundus therequires
images eyes. It very
causes affliction
skilled to the veins
practitioners due toin the
the
eyes causing
minute natureraptures on the
of features of retina which canwhich
the anomalies impactcanvision.
even If thetocondition
lead is failed
misdiagnosis andtoisbetime
detected at an early
consuming. stage, itdevising
Therefore, can lead an to
complete vision loss. The conventional diagnosing cycle of DR using fundus images requires very skilled
automated method for the diagnosis of DR can assist individuals with diabetes to identify symptoms of DR at a very early stage.practitioners due to the
minute nature also
This research of features of the anomalies
encompasses which can
the classification evendetected
of the lead to images
misdiagnosis
into itsand is time consuming.
corresponding 3 stagesTherefore,
namely No devising
Diabetic an
automated
Retinopathymethod for the
(No DR), diagnosis of DRDiabetic
Non-Proliferative can assist individuals(NPDR)
Retinopathy with diabetes to identify symptoms
and Proliferative of DR at a very
Diabetic Retinopathy early
(PDR) stage.
that can
This research
greatly also encompasses
aid in monitoring the classification
the dynamics of the key of the detected
features includingimages intohemorrhages,
lesions, its corresponding 3 stages
and density of namely No Diabetic
blood vessels. Deep
Retinopathy (No DR), Non-Proliferative Diabetic Retinopathy (NPDR) and Proliferative Diabetic Retinopathy
learning algorithms has gotten to become a well-known method that can accomplish a wide variety of classification tasks. However, (PDR) that can
greatly
most of aid in methods
these monitoring
are the
onlydynamics
efficient of the key features
in classifying including
the various lesions,
stages of DRhemorrhages, and density
with low accuracies of blood
notably, for thevessels. Deep
early stages.
learning
The algorithms
devised has in
algorithm gotten to becomeuses
this research a well-known methodprocessing
enriched images that can accomplish
techniques,a wide variety
automatic of classificationtuning
hyperparameter tasks.and
However,
neural
most of these
network methods
training are only
strategies efficient
to provide morein classifying
emphasis on thethe
various
minutestages of DR
features forwith lowprediction.
better accuraciesThe
notably, for the
algorithm early
was testedstages.
and
The devised algorithm in this research uses enriched images processing techniques, automatic hyperparameter
compared with modified Resnet50, VGG16, Mobilenetv2, Inceptionv3 and InceptionResnetv2 which gave a classification accuracy tuning and neural
network
of 94.7%,training
86.1%,strategies
85.8%, 85.3%to provide more
and 87% emphasis with
respectively on the minute features
corresponding for better
detection prediction.
accuracies The algorithm
of 99.8%, 94%, 94.2%, was 94.9%
tested and
and
compared
98.2% with modified
respectively on a Resnet50,
test set of VGG16,
508 images.Mobilenetv2,
Using theInceptionv3 and InceptionResnetv2
proposed algorithm, whichResnet50
the results indicate gave a classification
based network accuracy
gave
of 94.7%,performance
superior 86.1%, 85.8%, 85.3%
for both and 87%and
detection respectively withtasks.
classification corresponding detection accuracies of 99.8%, 94%, 94.2%, 94.9% and
98.2% respectively on a test set of 508 images. Using the proposed algorithm, the results indicate Resnet50 based network gave
superior
© 2023 performance
The Authors. forPublished
both detection and classification
by ELSEVIER B.V. tasks.
This is an
© 2023 Theopen access
Authors. article by
Published under the CC
Elsevier B.V.BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
© 2023 Theunder
Peer-review Authors. Published
responsibility of by
theELSEVIER
scientific B.V.
committee of the International Conference on Machine Learning and Data
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an
Engineering open access article under the CC BY-NC-ND license
Peer-review under responsibility of the scientific committee of the (https://creativecommons.org/licenses/by-nc-nd/4.0)
International Conference on Machine Learning and Data Engineering
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and Data
Engineering
Keywords: Diabetic Retinopathy; Convolutional Neural Network; Non-Proliferative Diabetic Retinopathy; Proliferative Diabetic Retinopathy

Keywords: Diabetic Retinopathy; Convolutional Neural Network; Non-Proliferative Diabetic Retinopathy; Proliferative Diabetic Retinopathy
* Corresponding author. Tel.: +91-9447205116
E-mail address: [email protected]
* Corresponding author. Tel.: +91-9447205116
E-mail address: [email protected]

1877-0509 © 2023 The Authors. Published by ELSEVIER B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
1877-0509
Peer-review©under
2023 The Authors. Published
responsibility by ELSEVIER
of the scientific committeeB.V.
of the International Conference on Machine Learning and Data Engineering
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and Data Engineering

1877-0509 © 2023 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and
Data Engineering
10.1016/j.procs.2023.01.085
1056 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066
2 Athira / Procedia Computer Science 00 (2023) 000–000

1. Introduction

Diabetic Retinopathy is an extreme complication of diabetes mellitus which even in the current days remains as
the leading cause of blindness globally [8]. Countries of huge population like China, Indonesia, India, and Bangladesh
shares 45% of the global counts in diabetes [1]. As the numbers are anticipated to move up, an automatic approach
for clinical diagnosis will be of much help. The primary motivation of this research is to come up with an application
for state of the art deep learning algorithms in the medical image analysis and diagnosis to aid people with long term
diabetes to identity and take precautionary actions against its progression. An automatic system would help people
with diabetes to identify the signs of the disease at an earlier stage which can highly reduce the clinical load on retina
experts and help monitor the dynamics of the lesions. Thus, the research on automated diagnosis of DR becomes more
and more crucial in the past few years. Through this research we propose an efficient algorithm for detection and a 3-
stage severity grading of DR. While most research do not give much emphasis on blood vessels the proposed approach
gives more importance to feature enhancement techniques that enhance blood vessels treating it as an important
feature. This research has also eliminated the need for manual hyperparameter tuning by incorporating autotuners to
find best possible parameters from a large search space and to dynamically add extra layers during runtime. In this
research a Convolutional Neural Networks approach is proposed to automate the method of DR screening using
preprocessed fundus retinal image as input.

(a) (b) (c)

Fig. 1. (a) Normal image (b) Red lesion (c) Exudate

Classes in DR classification include No DR, NPDR and PDR [7]:

• No Diabetic Retinopathy: At this stage no signs of DR can be seen at this stage.
• Non-Proliferative Diabetic Retinopathy is the earliest stage in DR where micro aneurysms such as red
lesions start to occur. The potential of blood transportation due to blood vessel distortion and swelling
decreases as the disease progresses. NPDR is characterized by the presence of minute red lesion in a scale
which is hard to be identified on visual inspection as depicted in Fig1b making it difficult to detect the
disease at this stage. Vision is not usually affected at this stage although there are higher risks for
developing eyesight issues in the future. If the disease is identified at this stage oral treatment can help
reverse the effect.
• Proliferative Diabetic Retinopathy is a highly progressed stage of DR, when the retinal blood vessels cause
activates proliferation of the newly formed blood vessels emerging inside in the vitreous gel in the eye.
Exudates becomes more prevalent at this stage. Laser treatment may have to be prescribed to avoid
complete vision loss. The slight yellow patches depicted inside the localized region in Fig1c shows an
example of exudates.

Each stage has its characteristics and numerous properties which in certain cases are hard for doctors to possibly
take note of, thereby increasing the chances of incorrect diagnosis. This led to the need for devising an automated
approach for DR detection.
Object detection and classification utilizing different machine learning procedures have been a vital focal point of
the exploration for the research society [2]-[3]. Particularly with the approach of CNN, different models have been
proposed to achieve the errands of computer vision (CV), speech recognition, natural language processing (NLP and
 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066 1057
Athira/ Procedia Computer Science 00 (2023) 000–000 3

advanced mechanics, [4]. There are also different instances of deep learning (DL) use in biomedical applications
[5],[18]-[23],[27],[28]. In this research, we have presented a strategy that covers the data preparation, recognition, and
classification of DR from fundus images. The primary focus by incorporating deep learning techniques in image
classification is to introduce efficient algorithms to replace the dependence hand coded key points for classifying DR
images.
Acharya et al. proposed a novel automated technique for identifying the DR stages using simple image processing
techniques. The algorithm gave an average accuracy, specificity and sensitivity as 85%, 86% and 82% respectively.
This result could be increased by the usage of more efficient feature enhancement techniques [9]. Sudeshna et al. has
done considerable work in extracting out blood vessel and removing optic disc, which in most cases may result in
creation of residues leading to a false detection. The research discussed techniques of morphological closing and
opening for the extracting out natural structures in the retina. But the similarity in the nature of color characteristics
for normal and abnormal features of in the retina can sometimes result in residues [10].
Feature extraction is one important step in any kind of disease detection. With the use of CNNs manual feature
extraction becomes no longer necessary, a single CNN network can automatically extract features and do classification.
The findings of several other works related to this research is consolidated in table 1.

Table 1. Summary of related works.

Sl No Authors Name Method Observations

1 Aacharya et.al, 2008 [9] Proposed automated technique to identify DR Simple image processing gives moderate
stages using simple image-processing and data- accuracy.
mining techniques using higher order spectra.
2 Doshi et.al, 2016 [16] GPU accelerated deep CNN based algorithm. Image processing techniques can be used for
decreasing overfitting through increasing
complexity of the data.
3 Wang et.al, 2018 [13] Three convolutional neural networks were Paper shows InceptionNet achieved the highest
separately analyzed. accuracy.
4 Sudeshna et.al, 2017 [10] Extracted and removed all features that are The approach resulted in enhanced the
possible candidates for false negatives. performance compared to existing approaches but
performs slightly weaker in detecting red lesions.
5 Wen et.al, 2020 [14] Designed a novel Resnet-50 based TCNN for Explored the capabilities of Resnet based model.
fault diagnosis.
6 Alban et.al, 2016 [15] Implemented DR classification using Google Net, The model performs well in comparison to
AlexNet and Baseline (Custom Architecture). human evaluation metrics and the performance of
weak learners can be boosted with the approach
of ensemble learning.

The rest of the paper is organized into various sections which mainly focus on methodology, results discussion,
conclusion and future work. The major steps in the implementation of the proposed algorithm starting from details
of the dataset to the training strategies are included in section 2. Section 3 covers the experimental setup and result
evaluations. Section 4 is a brief discussion of the paper. Finally, section 7 summarizes the conclusions and directions
for future enhancements.

2. Methodology

The proposed algorithm in this work utilizes enriched preprocessing and highly optimized parameter tuning for
modelling and training to ensure the best possible results. The whole setup can be broken down into preprocessing,
hyperparameter optimization, hybrid network training and finally analysis of the classification results from various
networks. The basic blocks of the proposed algorithm is summarized in figure 2. The dataset contains the images of
each of the 3 stages in DR. Image resizing, grey scaling, gaussian blurring and circular cropping were the image
preprocessing steps used to enhance the important features in the images. Preprocessing was followed by
hyperparameter tuning for each selected base network which included training parameters as well as model building
1058 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066
4 Athira / Procedia Computer Science 00 (2023) 000–000

parameter like addon layers. Instead of using a conventional manual tuning approach for hyperparameter optimization,
we implemented an hyperparameter optimizer using SLSQP to automize the whole process as well as to make sure to
always get the best set of hyperparameters from a search space. The best set of parameters for each of the 5 models
were used for the final model building and training. The classification results of each of the 5 models are then analyzed
and compared. The whole research was done on a Nvidia Tesla P100 GPU with a maximum RAM usage of 24 GB as
high RAM usage was necessary during preprocessing and auto hyperparameter tuning to avoid crashes and to increase
the speed of execution.

Fig. 2. Schematic Diagram of the Proposed Approach

2.1. Dataset

The dataset used is a subset of the Asia Pacific Tele Ophthalmology Society. The complete dataset consists of
18590 fundus images. The dataset is relatively inhomogeneous in terms of quality as these images were not obtained
in a controlled lab setup and used different cameras to capture resulting in different image resolutions [14]. To
overcome these challenges, the images in APTOS dataset were screened additionally by ophthalmologist whenever
required for confirmation. Therefore, subset of 3386 images were used for the train and test datasets and is reported
in this paper. 85% of the dataset was utilized for training, the rest 15% was utilized for testing. The classes in the data
includes No DR (class 0), NPDR (class 1) and PDR (class 2).

2.2. Preprocessing

Preprocessing techniques are chosen to provide more focus in enhancing the features like blood vessel, red lesions
and exudates. Due to the huge amount of data summing close to 10 GB, to reduce the time for preprocessing thread
pooling was used. Thus, the whole process of preprocessing was divided into 4 different threads and ran on 4 different
cores concurrently.

2.2.1. Resizing the images

As the data was collected from various sources at different time period, each image data had differing image size
and storage memory. Therefore, rescaling all the images in the entire dataset to a uniform image size was essential for
 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066 1059
Athira/ Procedia Computer Science 00 (2023) 000–000 5

the better computation capability. Figure 1 shows the visualization of 3 random images, chosen from the three
different classes. It was observed that the features such as red lesions and exudates were not clearly visible, especially
in the images which lacked proper lighting.

2.2.2. Convert to grey scale

Creating a better model, requires training the model on a uniform well defined dataset. The first technique used for
feature enhancement was converting the color images to grey scale images. The features such as exudates (yellow
spots) was portrayed very well, whereas the feature such as red lesions were barely visible. Training the model on
such dataset will result in very high bias and high training error. A far better result was observed using advanced
feature enhancement techniques such as gaussian blur and circular cropping.

2.2.3. Gaussian Blur

It is an image processing technique used to lessen the amount of noise in an image. An image, which is a huge two-
dimensional matrix of numbers representing the color of the pixel is passed through the gaussian filter. It is usually
accomplished by convolving pixel array with the gaussian filter. A 2-dimensional gaussian filter can generally be
expressed as:

𝒙𝒙𝟐𝟐 +𝒚𝒚𝟐𝟐
𝟏𝟏 −( )
𝑮𝑮𝟐𝟐𝑫𝑫 (𝒙𝒙, 𝒚𝒚, 𝝈𝝈) = 𝒆𝒆 𝟐𝟐𝝈𝝈𝟐𝟐 (1)
𝟐𝟐𝝅𝝅𝝈𝝈𝟐𝟐

Here x and y are the location indices and standard deviation of the distribution being denoted by σ. The idea of σ
controls the fluctuation around a mean of the gaussian distribution, which decides the degree of the obscuring impact
around a pixel [11]. A wide range of combinations of these parameters were experimented to identify the parameter
set with the best result. The gaussian blur utilized that gave the best blood vessel and red lesion enhancement result
had a sigma x which is the gaussian kernel standard deviation in x direction as 30. After applying gaussian filter, the
areas that contained more information were extracted. To sharpen the edges of the image obtained after gaussian blur.
A weighted masking is used to sharpen the edges of the image obtained after gaussian blur. The first image passed is
the resized image with alpha weighted as 5 and the second image passed is the image after applying gaussian blur
with beta weighted as -5. The gamma value is specified as 128. Images with both features of exudates and red lesions
and more importantly blood vessel were enhanced more than the original images were obtained as shown in the third
step of figure 3.

2.2.4. Circular cropping

To give more focus to the part containing Blur the fundus image of the eye, a circular cropping is supplied to all
the images to discard nonessential background hence making a uniform training dataset. This helps to fit the images

Fig. 3. Outputs of various stages in preprocessing

1060 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066
6 Athira / Procedia Computer Science 00 (2023) 000–000

to the new image frame secured after removing the background through the circular cropping technique as shown in
the figure 3. In order to apply circular cropping on a color image, the image is first converted to grey scale and a mask
is created which would pass only those pixel values of the image, which are greater than the specified tolerance value.
In cases where the image is too dark, we return the original image to avoid the cropping of the entire image. In the
other case, the boolean array obtained after applying the masking can be used to index the image for extracting required
bounding boxes making use of broadcasted indexing and a bitwise and operation is done between this image and the
dummy circle created for doing the circular cropping.

2.3. Visualizing Features

Feature space visualization was done to identify the degree of model complexity to choose the right model for the
classification task. A non-linear technique of feature visualization called t-Distributed Stochastic Neighbor
Embedding (t-SNE) was for information investigation and imagining high-dimensional information in an
unsupervised manner in this study. Unlike PCA this technique is more than just a mathematical procedure but is a
probabilistic technique. t-SNE limits the difference between two dispersions: one that that estimates pairwise
likenesses of the input objects and the other that estimates the pairwise likenesses of the low-dimensional data in the
embedding. [12]. Plotting t-SNE gave us an idea on the level of difficulty in separating the various classes in data.
From the figure 4 it can be noted that the classes within DR specifically, classes 0 and 2 have many overlapping
features hence difficult to distinguish. t-SNE plots proves that even though there are principal features that are
comparatively easy separable in class 0 there are a group of other features in class 0, 1 and 2 which are highly
convoluted and hard to separate out. The main feature of t-SNE is that it tries to deconvolute relationships between
neighbors in high-dimensional data giving us a better understanding of the separability of the classes and is a reducing
the dimensions for visualization.

Fig. 4. 2-Dimensional visualization of data with t-SNE

2.4. Training

Modified Resnet50, InceptionResnetv2, Mobilenetv2, Inceptionv3 and VGG16 were used for training. The
technique of transfer learning was employed to get the maximum out of it. After the elaborate training processes, it
was observed that Resnet based networks viz. Resnet50 and InceptionResnetv2 gave better performances compared
to the non-Resnet based networks. Resnet50 modified network gave the maximum accuracy among all the tested
networks. ResNet50 is a convolutional neural network with 50 layers. The pretrained variant of the network on
ImageNet data is used in the in this study [1]. Resnet is characterized by the presence of skip connections that greatly
reduces the effect of vanishing and exploding gradients. Resnet50, Mobilenetv2 and InceptionResnetv2 are the other
networks that uses skip connections in the basic blocks a sample of which is depicted in figure 5.
To the base Resnet model an additional block of 4 extra layers were added to get the final model for training. A
dropout of 0.3 was used for transfer learning. The final dense activation layer was of sigmoid with 3 outputs for the 3
different classes. The performance of the deep learning models is extremely sensitive to altering hyperparameters
 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066 1061
Athira/ Procedia Computer Science 00 (2023) 000–000 7

hence we created an hyperparameter optimizer module to test and tune all possible combinations of hyperparameters
from its possible sample space. Hence the best hyperparameters were found using auto tuners with a well-defined
search space, The auto-tuner used for the experimentation was a ray tuner. A tuner supports numerous search
algorithms viz PBT, Random Search, ASHA, Bayesian Optimization, Hyper Band, Median Stopping, BOHB. The
best hyperparameters were chosen for training.

Table 1. Addon layers in each model

Sl No Resnet 50 Inception Resnet v2 Mobilenet v2

1 Dense + activation = ‘relu’ Global Average Pooling2D Dense Layer
2 Dropout [ratio - 0.3] Dense + activation = ‘relu’ activation = ‘sigmoid’
3 Dense + activation = ‘relu’ Dense + activation = ‘sigmoid’
4 Dense + activation = ‘sigmoid’ .

Fig. 5. Basic building blocks of chosen base networks with skip connections

The automatic hyperparameter tuner picks up the best set of hyperparameters from a well-defined search space
thereby making sure to get the best combination of hyperparameters. The parameters that were allowed to be chosen
by the tuner include batch size, learning rate, number of addon layers, inclusion or exclusion of batch normalization
layers, number of epochs to train the choice of enforcing positive bias and weight constraints.

Fig. 6. Hyperparameter Tuning Process

8 Athira / Procedia Computer Science 00 (2023) 000–000
1062 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066

Figure. 6 shows the autotuner block in the workflow. Figure 7 shows the basic algorithm used to implement the
hyperparameter tuner [17][24].

Algorithm 1: Basic algorithm of hyperparameter tuning

Input (Hyperparameter space θ, Search algorithm S, Evaluation function E (θ), max number of evaluations n max)
Select initial set of hyperparameters θ 0 Є θ
Evaluate initial score yo = E (θ 0)
Set θ * = θ 0 and y* = E (θ 0)
For n=2…nmax
Select new set of hyperparameter θ new from defined space using defined Search algorithm S
Calculate Enew with new parameters to get ynew= E (θ new)
If ynew < Threshold error (extremely small value)
Assign θ * = θ new and y* = ynew
Return θ *= θ new and y* = ynew

Fig. 7. Algorithm of hyperparameter tuning

It is prescribed to utilize batch sizes with powers of 2 since it fits with the memory of the CPU. For transfer learning,
an extremely low learning rate is favored with the goal that it doesn't change a lot of what is recently learnt by the
network. The quantity of addon layers relies on the amount of learnt patterns that can be transferred from the layers
of the pre-trained model. Involving every one of the layers for transfer learning, a basic flatten layer and dense layer
with softmax is sufficient however since we integrated the feature extraction it required more layers toward the end.
A few Optimizers were experimented including stochastic gradient descent (SGD) and root mean squared propagation
(RMSprop). SGD with an exceptionally low learning rate expected more iterations to finish training the model with a
reasonable amount of pattern being learnt. Thus, RMSprop was utilized to come by the expected outcome. Batch
normalization and dropout layers are placed in between the dense layers to avoid any chances of overfitting. Training
was done in 4 steps with 50 epochs in each step. The initial 2 epochs were trained on a very low learning rate as a
warmup. The whole network was trained 200 epochs excluding the epochs for warmup with a slightly higher learning
rate of 0.001 finally to yield a training accuracy of 98.9%.

Table 2. Autotuner model building and training parameter space

Sl No Hyper parameter Optimizer search space

1 Epochs [50, 100, 150, 200, 500, 1000]
2 Batch Size [4, 8, 16, 32, 64]
3 Learning Rate [10−1, 10−2, 10−3, 10−4, 10−5, 10−6]
4 Warmup Epochs [1, 2, 5, 10, 15]
5 Image Height, Width [126, 224, 512]
6 Early stopping Patience [5, 8, 10, 12, 15]
7 RLprop Patience [2, 3, 4, 5, 8, 10]
8 Decay Drop [0.1, 0.2, 0.3, 0.4, 0.5]
9 Addon Dense layers [1, 2, 3, 4, 5, 6]
10 Addon Dropout layers [1, 2, 3, 4, 5, 6]
11 Dropout ratio [0, 0.1, 0.2, 0.3, 0.4, 0.5]
 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066 1063
Athira/ Procedia Computer Science 00 (2023) 000–000 9

There are four main approaches used during training - A warmup phase, training with higher learning rate, reducing
learning rate on plateau and early stopping.

2.4.1. Warmup phase

Warm up step is used for reducing the learning rate to decrease the effect of deviating the model from learning the
immediate new data set to which it is exposed. It helps lower the impact of the primacy effect due to the early training
examples. In the absence of a warmup phase, it might be required run extra epochs to get the desired convergence.
The first 2 epochs were allocated for warmup training with a learning rate of 0.0001. It also helps fight early overfitting
[6].

2.4.2. Training with higher training rate

After training 2 epochs with lower learning rates the rest of the 200 epochs are trained with higher learning rates of
0.001 after the warmup phase for the Resnet50 based network. Hyperparameter tuners were used to identify the
optimal parameters from a large search space to ensure the best accuracy.

2.4.3. Reducing Learning rate on plateau

This approach was for reducing the learning rate when the matric stops improving. We monitored the validation
loss of consecutive 3 epochs for creating a decay drop in case of not much improvement in the metric measured.

2.4.4. Early stopping

Early stopping is an effective strategy to avoid overtraining. The training process is halted as soon as the
performance on the current validation cycle diminishes comparing the last validation data performance. Using the
above stated strategies, Resnet50 based model was able to attain a training accuracy of 98.9%. with a validation
accuracy of 90.88%.

3. Results
The experimentation was done in two different approaches to separately evaluate all the 5 models for their disease
detection and disease classification abilities. 508 images were chosen as the test set for both the testing. Initially a 3-
class testing with classes No DR, NPDR and PDR was done to assess the proficiency of the model on grading the
severity of the disease. A second level of testing was done to evaluate the disease detection capability of the models.
To test the detection capability all class 0 images were as taken as the non-DR class and data from class 1 and 2 as
DR class. The data was tested on modified Resnet50, InceptionResnetv2, Mobilenetv2, Inceptionv3 and VGG16.

Fig. 8. Analysis of overall accuracy on experimented networks

10 Athira / Procedia Computer Science 00 (2023) 000–000
1064 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066

A total of 4 evaluation matrices were used to assessing the model’s performance on unseen data which includes,
accuracy, precision, recall, f1 score and kappa score. These results are consolidated in table 3. As this is a multi-class
problem the overall values of the evaluation matrices are found in three ways as micro, macro and weighted for the
best network is consolidated in table 4. The disease detection testing showed an accuracy of 99.8% and disease
classification 94.7% on Resnet50 based network as shown in figure 8.

Table 3. Overall Comparison table for detection and classification based on accuracy and kappa score
Evaluation Matric Resnet 50 VGG − 16 MobilenetV2 InceptionResnetV2 Inception V3
Classification Accuracy 0.947 0.861 0.858 0.870 0.852
Kappa Score 0.884 0.713 0.739 0.929 0.724
Detection Accuracy 0.998 0.942 0.94 0.982 0.949

Table 4. Detailed analysis of Resnet 50 based network

Detection Classification
Precision Recall F1-score Precision Recall F1-score
Micro 0.998 0.998 0.998 0.94 0.94 0.94
Macro 0.999 0.997 0.998 0.92 0.81 0.83
Weighted 0.998 0.998 0.998 0.94 0.94 0.93

The analysis of results with the same algorithm done for 4 other major networks is consolidated in figure 8 and
figure 9. Using the above algorithm Resnet50 based network is tested to show an overall higher performance in terms
of all the evaluated parameters taking all the 3 classes together. It can also be noted that InceptionResnetv2 showed
higher performance considering precision of class 2 alone thereby opening possibilities of even higher performance
by combining the 2 networks. This result can be attributed to the architecture of Resnet based networks. The presence
of skip connections helps propagate the importance of minute features without being degraded by activation functions
or other mathematical computations.

Fig. 9. Classwise Analysis of performance parameters of experimented with major networks

Rao at al. experimented with 5 different state-of-the-art deep learning models with Adam and Stochastic Gradient
Descent (SGD) on 2 different input sizes viz 224x224 and 512x512 to give promising results as depicted in Table 5
and 6. On comparison with the work done by Rao at el. For the networks common in both the studies for detection
both gave higher performance with the Resnet50 network. With our approach Resnet50 and inception v3 gave
 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066 1065
Athira/ Procedia Computer Science 00 (2023) 000–000 11

significantly higher performance and VGG16 gave extremely closer performance to that of Rao at el. Analysis of the
classification results of both the research shows that Rao at el. achieved a maximum accuracy of 88.1% on
InceptionResnetv2, and our work achieved 94.7% with the much lighter Resnet50 based network.

Table 5. Detection performance comparison of proposed approach with prior works

Proposed Rao et al [26]
Resnet 50 VGG 16 Inceptionv3 Resnet 50 VGG 16 Inceptionv3
Accuracy 0.998 0.942 0.949 0.966 0.95 0.896
Precision 0.99 0.94 0.95 0.97 0.95 0.89
Recall 0.99 0.95 0.95 0.97 0.95 0.89
F1 Score 0.99 0.95 0.95 0.965 0.950 0.892

Table 6. Classification performance comparison of proposed approach with prior works

Proposed Rao et al [26]
Resnet50 VGG16 Inceptionv3 Inception Resnet50 VGG16 Inceptionv3 Inception
Resnetv2 Resnetv2
Accuracy 0.947 0.861 0.852 0.87 0.785 0.747 0.812 0.881
Precision 0.94 0.83 0.84 0.86 0.79 0.76 0.83 0.88
Recall 0.94 0.85 0.85 0.87 0.78 0.76 0.81 0.88
F1 Score 0.93 0.83 0.84 0.87 0.778 0.756 0.805 0.882

4. Discussion

This research proved that machine learning algorithms such as neural networks do have high future scope in
detecting abnormalities from medical images. Research have proactively demonstrated the proficiency of CNN of
CNN technique in tasks like object detection. With the results from this research, it is evident that the scope of neural
networks is not just confined in finding object of considerable sizes and definite shapes, but neural networks can also
be successfully used for classifying objects with very tiny features with irregular shapes and sizes. The test accuracy
of 99.8% for detection and 94.7% on classification obtained using Resnet based networks proves that with the right
kind of image enhancement in preprocessing CNN can be found to be highly proficient for tasks like abnormality
detection and classification from medical images.

5. Conclusion and Future Work

From the research it can be observed that networks with skip connections performed well in classifying the stages
when the feature are extremely small. Resnet50 based network topped the overall performance matrices for both
detection and classification proving that with the right set of hyperparameters and preprocessing light weight skip
connected networks can be efficiently utilized for DR detection and classification. The proposed system can greatly
reduce the time of diagnosis and thus improve the quality of life of diabetic patients. Improvements in this works can
be made in a several viewpoints. The 3-level classification can be extended to a more elaborate 5 class problem making
use of more advanced features. An ensemble approach can be implemented to boost the overall proficiency of the
algorithm. The power of CNN can be leveraged for optic disc removal and blood vessel extraction for enhancing
classification performance without causing residues. CNN based image semantic segmentation for cotton wool
natured exudates which are hard to detect using normal image processing methods could likewise be examined to
carry out fine-grained DR classification.
1066 Athira T R et al. / Procedia Computer Science 218 (2023) 1055–1066
12 Athira / Procedia Computer Science 00 (2023) 000–000

References

[1] D. S. Fong, F. L. Ferris III, M. D. Davis, E. Y. Chew, E. T. D. R. S. R. Group. (1999) “Causes of severe visual loss in the early treatment
diabetic retinopathy study: Etdrs report no. 24”, American journal of ophthalmology 127 (2): 137–14
[2] M. Jones, P. Viola. (2003) “Fast multi-view face detection.”, Mitsubishi Electric Research Lab TR-20003-96 3 (14): 2.
[3] W. Cheung, G. Hamarneh. (2009) “n-sift n-dimensional scale invariant feature transform”, IEEE Transactions on Image Processing 18 (9):
2012–2021.
[4] A. Krizhevsky, I. Sutskever, G. E. Hinton. (2012) “Imagenet classification with deep convolutional neural networks”, Advances in neural
information processing systems 25.
[5] L. Yang, Y. Zhang, J. Chen, S. Zhang, D. Z. Chen. (2017) “Suggestive annotation: A deep active learning framework for biomedical image
segmentation”, in: International conference on medical image computing and computer assisted intervention, Springer: 399–407
[6] F. He, T. Liu, D. Tao. (2019) “Control batch size and learning rate to generalize well: Theoretical and empirical evidence”, Advances in Neural
Information Processing Systems 32.
[7] W. R. Memon, B. Lal, A. A. Sahto. (2017) “Diabetic retinopathy; Frequency at level of hba1c greater than 6.5%”, The Professional Medical
Journal 24 (02): 234–238.
[8] A. W. Stitt, T. M. Curtis, M. Chen, R. J. Medina, G. J. McKay, A. Jenkins, T. A. Gardiner, T. J. Lyons, H.-P. Hammes, R. Simo. (2016) “The
progress in understanding and treatment of diabetic retinopathy”, Progress in retinal and eye research 51: 156–186.
[9] R. Acharya U, C. K. Chua, E. Ng, W. Yu, C. Chee. (2008) “Application of higher order spectra for the identification of diabetes retinopathy
stages”, Journal of medical systems 32 (6): 481–488.
[10] S. S. Kar, S. P. Maity. (2017) “Automatic detection of retinal lesions for screening of diabetic retinopathy”, IEEE Transactions on Biomedical
Engineering 65 (3): 608–618.
[11] S. Misra, Y. Wu. 2020 “Chapter 10 - machine learning assisted segmentation of scanning electron microscopy images of organic-rich shales
with feature extraction and feature ranking”, in: S. Misra, H. Li, J. He (Eds.), Machine Learning for Subsurface Characterization, Gulf
Professional Publishing: 289–314.
[12] L. Van der Maaten, G. Hinton. (2008) “Visualizing data using t-sne”, Journal of machine learning research 9 (11): 2579-2605
[13] X. Wang, Y. Lu, Y. Wang, W.B. Chen. (2018) “Diabetic retinopathy stage classification using convolutional neural networks”, in: 2018 IEEE
International Conference on Information Reuse and Integration (IRI), IEEE: 465–471.
[14] L. Wen, X. Li, L. Gao. (2020) “A transfer convolutional neural network for fault diagnosis based on resnet-50”, Neural Computing and
Applications 32 (10): 6111–6124.
[15] M. Alban, T. Gilligan. (2016) “Automated detection of diabetic retinopathy using fluorescein angiography photographs”, Report of standford
education.
[16] D. Doshi, A. Shenoy, D. Sidhpura, P. Gharpure. (2016) “Diabetic retinopathy detection using deep convolutional neural networks”, in: 2016
International Conference on Computing, Analytics and Security Trends (CAST) IEEE: 261–266.
[17] B. G. Galuzzi, I. Giordani, A. Candelieri, R. Perego, F. Archetti, (2020) “Hyperparameter optimization for recommender systems through
bayesian optimization”, Computational Management Science 17 (4): 495–515.
[18] R. Priya, V. Sreelekshmi, J. J. Nair, G. Gopakumar. (2021) “Breast mass classification using classic neural network architecture and support
vector machine”, in: Advances in Computing and Network Communications, Springer: 435–448.
[19] R. Dhanya, I. R. Paul, S. S. Akula, M. Sivakumar, J. J. Nair. (2020) “F-test feature selection in stacking ensemble model for breast cancer
prediction”, Procedia Computer Science 171: 1561–1570.
[20] R. Dhanya, I. R. Paul, S. S. Akula, M. Sivakumar, J. J. Nair. (2019) “A comparative study for breast cancer prediction using machine learning
and feature selection”, in: 2019 International Conference on Intelligent Computing and Control Systems (ICCS), IEEE: 1049–1055.
[21] A. Sreekumar, K. Nair, S. Sudheer, H. G. Nayar, J. J. Nair. (2020) “Malignant lung nodule detection using deep learning”, 2020 International
Conference on Communication and Signal Processing (ICCSP): 0209–0212.
[22] L. S. Nair, R. Prabhu, G. Sugathan, K. V. Gireesh, A. S. Nair. (2021) “Mitotic nuclei detection in breast histopathology images using yolov4”,
in: 2021 12th International Conference on Computing Communication and Networking Technologies (ICCCNT), IEEE: 1–5.
[23] K. Vaishnavi, M. A. Ramadas, N. Chanalya, A. Manoj, J. J. Nair. (2021), “Deep learning approaches for detection of covid-19 using chest x-
ray images”, in: 2021 Fourth International Conference on Electrical, Computer and Communication Technologies (ICECCT), IEEE: 1–8.
[24] J. Bergstra, D. Yamins, D. Cox. (2013) “Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision
architectures”, in: International conference on machine learning, PMLR: 115–123.
[25] Diabetic retinopathy detection: identify signs of diabetic retinopathy in eye images, https://www.kaggle.com/c/diabetic-retinopathydetection.
[26] Rao, Mihir, Michelle Zhu, and Tianyang Wang. (2020) “Conversion and implementation of state-of-the-art deep learning algorithms for the
classification of diabetic retinopathy”, arXiv preprint arXiv:2010.11692
[27] Singh, Vijendra, Vijayan K. Asari, and Rajkumar Rajasekaran. (2022) “A Deep Neural Network for Early Detection and Prediction of Chronic
Kidney Disease”, Diagnostics 12 (1): 116.
[28] Rastogi, Priyanka, Vijendra Singh, and Monika Yadav. (2018) “Deep learning and big data technologies in medical image analysis”, 2018
Fifth International Conference on Parallel, Distributed and Grid Computing (PDGC). IEEE: 60-63.