2023 International Conference on Advanced Computing Technologies and Applications (ICACTA)

Analysis of Transfers Learning Techniques

2023 International Conference on Advanced Computing Technologies and Applications (ICACTA) | 979-8-3503-4834-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICACTA58201.2023.10392325

for Early Detection and Grading of Diabetic

Retinopathy on Retinal Images
Santosh Kumar Sahu Ankush D. Sawarkar Ajay Kumar Sahu
Indian Institute Of Information Visvesvaraya National Institute Of Bajaj Institute Of Technology (BIT),
Technology, Nagpur, Maharashtra- Technology, Nagpur, India Wardha, Maharashtra, India
India ORCID : 0000-0001-7099-1987 [Link]@[Link]
ORCID : 0000-0003-4054-6078 ankush1sawarkar@[Link]
[Link]@[Link]
Akhil Anjikar Amol P Bhopale Jagdish Chakole
Jhulelal Institute of Technology (JIT) Indian Institute Of Information Indian Institute Of Information
Nagpur, Maharashtra, India Technology, Nagpur, Maharashtra-India Technology, Nagpur, Maharashtra-
akhil.anjikar09@[Link] amolpbhopale@[Link] India
jagdishchakole@[Link]

Abstract: Eye complications of diabetes include diabetic began to understand the specific mechanisms by which
retinopathy (DR). It's caused by retinal blood vessel diabetes damages the blood vessels in the retina, leading
alterations. Working-age individuals' major cause of to diabetic retinopathy. In the early 20th century, doctors
blindness is diabetic retinopathy. Early identification of began to recognize the link between diabetes and
diabetic retinopathy may prevent or postpone visual loss. blindness, with many patients with diabetes suffering
Artificial intelligence (AI) can use computational image
from vision loss due to diabetic retinopathy. However, at
analysis methods to automatically identify and categorize
diabetic retinopathy in retinal pictures. In this paper we the time, there were no effective treatments for the
study and analyze the different transfer learning models for condition. In the mid-20th century, researchers began to
early detection of diabetic retinopathy for retinal images. develop laser therapy as a treatment for diabetic
According to our findings, the accuracy of predictions retinopathy. In the late 1960s and early 1970s, a number
ranges from 74% to 81%. It is worth noting that AI-based of studies were published showing that laser therapy
systems for diabetic retinopathy detection are still in the could be used to reduce the risk of vision loss in patients
research phase, and more research is needed to evaluate with diabetic retinopathy.
their accuracy and effectiveness in real-world settings.

Keywords: Artificial intelligence, Deep Learning, Transfer

In the 1980s and 1990s, advances in imaging technology,
Learning, Diabetic Retinopathy, Multiclass Classification. such as the development of digital fundus cameras, made
it possible to detect diabetic retinopathy at an early stage,
I. INTRODUCTION before patients had significant vision loss. This has led
to the development of screening programs for diabetic
The history of diabetic retinopathy dates back to the early retinopathy, which have been shown to be effective in
19th century, when physicians first began to recognize reducing the risk of vision loss from the condition.
the association between diabetes and eye disease. More recently, with the development of advanced
However, it wasn't until the 20th century that doctors imaging techniques such as Optical Coherence

Class: 0- NO_DR Class: 1- Mild Class: 2- Moderate Class: 3- Severe Class: 4- Proliferate

Figure 1. Category of retinal images (1) No_DR (2) Mild (3) Moderate (4) Severe (5) Proliferate

979-8-3503-4834-7/23/$31.00 © 2023 IEEE

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on February 12,2025 at 08:17:31 UTC from IEEE Xplore. Restrictions apply.
Tomography (OCT) and AI methods, the early detection predictions and performance. To identify diabetic
and management of diabetic retinopathy have become retinopathy, updated DenseNet101 and ResNeXt are
more accurate, efficient and effective. Another approach is ensembled. ResNeXt improves ResNet models. Stacking
using the fundus images, an AI model trained to detect layers, split–transform–merge method, and a shortcut
retinopathy by analyzing the fundus images, can be used to between blocks are included in the model [10].
screen for the condition in a large population. The model
can be used to identify individuals who are at high risk of II. RELATED WORK
diabetic retinopathy and who should be referred for further
testing and treatment [1]–[4]. Following figures-1 shows Article provides a comprehensive evaluation of the
different category of retinal image. application of deep learning methods to the different
phases of the fundus based on images diabetic retinopathy
The majority of research in this area make use of fundus diagnosis pipeline. In addition, talk about various facets of
pictures, which are photographic recordings that detail the that pipeline, such as the commonly used datasets by the
current ophthalmic look of the retina in a person's eye. The research community, the preprocessing techniques used,
existence of DR symptoms in these fundus pictures may be and how these accelerate and improve the performance of
utilized to identify it by following numerous processes the models, and the creation of such deep learning models
such as the segmentation of retinal blood vessels, the for the diagnosis, grading, and localization of the disease
segmentation of lesions, and the identification of DR [2]. and its lesions. The authors elaborate on some models that
A wide range of research methodologies for the diagnosis have found practical use in actual clinical situations.
of DR at an early stage. In general, the findings of all Finally, we offer some key takeaways and suggest avenues
investigations point to the conclusion that the symptoms for further study.[11]. High-level characteristics from the
that appear during an early stage of retinopathy may final fully-connected layer of a Convolutional Neural
develop even in the absence of apparent clinical evidence Network (CNN) trained with transfer learning are used as
of retinopathy. The earliest recognition of DR is required input features for a Support Vector Machine (SVM)
since there is evidence that both functional and structural classifier. This technique uses CNN with fine-tuning to
damage has occurred. Although there are significant speed up the classification process.[12].
attempts being made to develop strategies for the early
identification of DR, none of the approaches that are now The bulk of this research will focus on evaluating current
being used have been shown to have a reliable predictive early DR diagnostic techniques for their strengths and
value [5]. They developed a compact using transfer weaknesses. Despite the prevalence of fluorescein
learning to identify the severity of DR and DME in both angiography images, early diagnosis can also be achieved
eyes at the same time. A data set consisting of 3028 images with color fundus medical imaging or visual lucidity
was used for the experiment's analysis, and five-fold cross- tomography angiography. In this analysis, only color
validation was performed on that set. As can be seen from fundus photographs were used. The early DR detection
our trial data, model has an average accuracy of 96.66%. methods discussed in this study can be placed into one of
[6]. Examines and analyses a number of different deep three broad categories: classical image processing,
neural networks in order to provide a grade to the illness traditional machine learning, and deep learning.[13]. The
according to severity levels such as mild, moderate, severe, purpose of this work is to compare and contrast several DR
and proliferative. On the Kaggle (APTOS) data set, pre- detection and classification systems based on machine
trained model such as VGG16, ResNet50, learning methods. Massive volumes of retina fundus and
and EfficientNetB5 achieve accuracies of 76.47%, 90.2%, thermal images from various public datasets are used for
and 97.2%, respectively [7]. "Average pooling" layer training and testing these systems. These methods have
outputs are used as features in the classifier created by the been shown to be effective in locating early indicators and
Support Vector Machine (SVM) with Error Correction grading the severity of a DR. ResNet50, a deep
Output Code. The recommended strategy's 80.1% severity convolutional neural network, was shown to be the most
class prediction accuracy outperforms the fine-tuned pre- effective algorithm for performance metrics across all
trained networks. This suggests that incorporating several studied systems.[14].
attributes from pre-trained networks might improve
recognition. [8]. Transfer learning approaches map picture In order to classify the degree of severity of disease in
labels to the proposed model's hyper parameters. Kaggle retinal fundus images, this study introduces a Frame-wise
provides retinal pictures and severity scales for training Severity Scale Classification Model (FSSCM) based on
and evaluating the model. Images range from normal to Transfer Learning enabled EfficientNet B3 and Fine
DR. The model detects and predicts DR from retinal Tuning enabled ResNet 101, respectively abbreviated as
pictures with 94.92% accuracy [9]. The automatic TL-EN3 and FT-RN 101.[15]
ensemble deep learning model detects and classifies DR.
III. METHODS
Putting together a deep learning model improves

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on February 12,2025 at 08:17:31 UTC from IEEE Xplore. Restrictions apply.
The publicly available dataset obtained from Kaggle’s Figure 3: Ten transfer learning models applied for grading of retinal
images.
Diabetic Retinopathy blindness Detection APTOS 2019
consist 3662 retinal images. Images are divided in to five
major categories. The image resolution is (224 x 224). Out of total 3662 images, we have used 90% images (3262)
These are No-DR, Mild, Moderate, Severe and Proliferate. for training and 10% images (367) for testing.
Table 1 shows number of images belongs to the particular After that we applied ten different transfers learning to
category (Table 1). training images and then tested the trained model to testing
images. The details of the analysis are given in
Table 1: Different category of images and their count Table 3. Artificial neural networks identify images using a
deep learning model. Image categorization deep learning
Class Image Category Count models often use multiple layers of artificial neurons.
0 No_DR 1805 These layers recognize and predict picture classification
characteristics. Convolutional layers identify edges and
1 Mild 370
other significant characteristics by applying filters to the
2 Moderate 999 picture. Subsequent layers normally link to all neurons in
3 Severe 193 the preceding layer. These layers collect form and texture
4 Proliferate 295 characteristics from the convolutional layer's output. The
Total 3662 output layer predicts an image label. This label is produced
Different categories of retinal images are shown in below from the last layer neurons' weighted values.
in graphical view. Here maximum images are belongs to
category of No_DR. (figure 2) A. Convolutional neural network (CNN): In the realm of
deep learning, convolutional neural networks (CNNs) are
widely used for various image and video processing
applications.

Figure 2: Image Category of different retinal images

Figure 4: Basic Architecture of CNN

The basic architecture of a CNN consists of several layers,

including:

• Input layer: The network receives raw picture data

here.
• CNN core: Convolutional layers. The picture is
filtered to extract features.
• Activation layers add non-linearity to the network.
ReLU is most popular.
• Pooling layers minimize feature map size.
• Fully linked layers produce final forecasts.
• The output layer makes image predictions.

B. Transfer Learning: Transfer learning uses a model

from one task to start a model for another. Transfer

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on February 12,2025 at 08:17:31 UTC from IEEE Xplore. Restrictions apply.
learning uses the model's first-task expertise to enhance its for image classification tasks. The Python-based Keras
second-task performance, even if the data and problem are package has multiple pre-trained transfer learning models.
different. Pre-trained (CNNs) as feature extractors for Some of the most commonly used transfer learning
image categorization are a frequent transfer learning methods in Keras include (Table 2):
example. Even with diverse data, these pre-trained CNNs,
trained on huge datasets like ImageNet, may be employed
Table 2: Some transfer learning models their memory size, top1 and top5 accuracy parameter and depth.

Size Top-1 Top-5

SN Model Parameters Depth
(MB) Accuracy Accuracy
1 MobileNet 16 70.40% 89.50% 4.3M 55
2 ResNet50V2 98 76.00% 93.00% 25.6M 103
3 Xception 88 79.00% 94.50% 22.9M 81
4 DenseNet201 80 77.30% 93.60% 20.2M 402
5 InceptionV3 92 77.90% 93.70% 23.9M 189
6 DenseNet169 57 76.20% 93.20% 14.3M 338
7 ResNet101V2 171 77.20% 93.80% 44.7M 205
8 ResNet152V2 232 78.00% 94.20% 60.4M 307
9 NASNetMobile 23 74.40% 91.90% 5.3M 389
10 InceptionResNetV2 215 80.30% 95.30% 55.9M 449

*The top-1 accuracy and top-5 accuracy measures how includes levels like "activation," "batch normalization,"
well the model did on the ImageNet evaluation dataset. and so on.
Depth refers to the network's topological depth. This
IV. RESULTS efficacy of each model. We performed experiment with
Ten distinct transfer learning models were tested across 100 epochs. Training and testing examples might affect
five distinct categories of retinal pictures. Training the final output. Results may vary according to training
accuracy, validation accuracy, testing accuracy, and testing examples.
precision, recall, and F1 score were used to evaluate the
Table 3: Experimental results of transfer learning models

Training Validation Testing F1-

Model Name Precision Recall
Accuracy Accuracy Accuracy Score
MobileNet 91.01 79.51 81.29 80.00 81.00 80.00
ResNet50V2 93.30 75.86 79.38 80.00 79.00 79.00
Xception 89.39 78.60 78.84 79.00 79.00 78.00
DenseNet201 89.46 76.77 77.20 80.00 77.00 78.00
InceptionV3 90.84 77.38 81.29 81.00 81.00 81.00
DenseNet169 88.45 79.51 80.20 77.00 80.00 78.00
ResNet101V2 94.82 76.47 78.84 79.00 79.00 78.00
ResNet152V2 94.78 76.16 79.11 78.00 79.00 78.00
NASNetMobile 85.55 74.64 77.20 76.00 77.00 76.00
InceptionResNetV2 83.49 77.68 78.02 73.00 78.00 75.00

V. DISCUSSION diabetic retinopathy of retinal images" provides a

detailed examination of the many available transfer
Diabetic retinopathy diagnosis and grading is a crucial learning methods for this specific application. We
duty in ophthalmology. Diabetic retinopathy diagnosis is evaluate the performance of various transfer learning
only one area where the accuracy of image classification techniques on retinal images and calculated their
tasks like transfer learning has shown great promise in accuracy, precision, recall and f1-score. We found that
recent years. The article "Analysis of various transfer MobileNet, InceptionV3 and DenseNet169 perform well
learning techniques for early detection and grading of among all.

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on February 12,2025 at 08:17:31 UTC from IEEE Xplore. Restrictions apply.
 Results of various transfer learning model
100
80
60
40
20
0

Training Accuracy Validation Accuracy Testing Accuracy

Precision Recall F1-Score

Figure 5: Ten transfer learning models applied for grading of retinal images and their training and testing accuracy

Models Name and Testing Accuracy

82
81
80
79
78 81.29 81.29
80.2
79.38 78.84 78.84 79.11
77
78.02
76 77.2 77.2
75

Figure 6: Models name and their testing accuracy

VI. CONCLUSION Transfer learning is a machine learning technique that

has proven to be effective in the classification of medical
Diabetes may lead to a number of serious complications, images, including retinal images. In this study, we have
the most serious of which is diabetic retinopathy, which applied ten models of transfer learning on images of
is the major cause of visual loss in diabetics. Diabetic diabetic retinopathy and analyzed their results. In
retinopathy must be diagnosed and treated as soon as conclusion, transfer learning
possible in order to avoid irreversible visual loss.
has the potential to revolutionize diabetic retinopathy REFERENCES
detection by providing an accurate and efficient way to
diagnose the disease in its early stages. This can lead to [1] R. Thanki, “Healthcare Analytics A deep neural network and
machine learning approach for retinal fundus image
better patient outcomes, improved vision, and a higher classification,” Healthc. Anal., vol. 3, no. January, p. 100140,
quality of life for individuals affected by diabetic 2023, doi: 10.1016/[Link].2023.100140.
retinopathy. [2] A. Sebastian, O. Elharrouss, and S. Al-maadeed, “diagnostics
A Survey on Deep-Learning-Based Diabetic Retinopathy
Classification,” pp. 1–22, 2023.
[3] A. Malhi, R. Grewal, and H. S. Pannu, Detection and diabetic

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on February 12,2025 at 08:17:31 UTC from IEEE Xplore. Restrictions apply.
 retinopathy grading using digital retinal images, no. Diagnostics, vol. 13, no. 1, pp. 1–14, 2023, doi:
0123456789. Springer Nature Singapore, 2023. doi: 10.3390/diagnostics13010124.
10.1007/s41315-022-00269-5. [11] N. Tsiknakis et al., “Deep learning for diabetic retinopathy
[4] S. Li, R. Zhao, H. Zou, and L. Guo, “Artificial intelligence for detection and classification based on fundus images: A
diabetic retinopathy,” Chin. Med. J. (Engl)., vol. 135, no. 3, pp. review,” Comput. Biol. Med., vol. 135, p. 104599, 2021, doi:
253–260, 2022, doi: 10.1097/CM9.0000000000001816. 10.1016/[Link].2021.104599.
[5] H. Safi, S. Safi, A. Hafezi-Moghadam, and H. Ahmadieh, [12] D. U. N. Qomariah, H. Tjandrasa, and C. Fatichah,
“Early detection of diabetic retinopathy,” Surv. Ophthalmol., “Classification of diabetic retinopathy and normal retinal
vol. 63, no. 5, pp. 601–608, 2018, doi: images using CNN and SVM,” Proc. 2019 Int. Conf. Inf.
10.1016/[Link].2018.04.003. Commun. Technol. Syst. ICTS 2019, pp. 152–157, 2019, doi:
[6] Zhenzhen lu, J. Miao, J. Dong, S. Zhu, X. Wang, and J. Feng, 10.1109/ICTS.2019.8850940.
“Automatic Classification of Retinal Diseases with Transfer [13] A. V. Kumar and A. S. Babu, “Diabetic Retinopathy Detection
Learning-Based Lightweight Convolutional Neural Network,” using Deep Learning Methodology,” 2022 IEEE 3rd Glob.
SSRN Electron. J., vol. 81, no. November 2022, p. 104365, Conf. Adv. Technol. GCAT 2022, 2022, doi:
2022, doi: 10.2139/ssrn.4183380. 10.1109/GCAT55367.2022.9971891.
[7] A. T. Nair, M. L. Anitha, and M. N. Arun Kumar, “Disease [14] D. A. Hasan, S. R. M. Zeebaree, M. A. M. Sadeeq, H. M.
Grading of Diabetic Retinopathy using Deep Learning Shukur, R. R. Zebari, and A. H. Alkhayyat, “Machine
Techniques,” Proc. - 6th Int. Conf. Comput. Methodol. Learning-based Diabetic Retinopathy Early Detection and
Commun. ICCMC 2022, no. Iccmc, pp. 1019–1024, 2022, doi: Classification Systems - A Survey,” 1st Babylon Int. Conf. Inf.
10.1109/ICCMC53470.2022.9754113. Technol. Sci. 2021, BICITS 2021, vol. 2021, no. Bicits, pp. 16–
[8] K. Ka, Y. Tiong, and F. H. Juwono, “Diabetic Retinopathy 21, 2021, doi: 10.1109/BICITS51482.2021.9509920.
Detection : Improving Accuracy Using Multiple Transfer [15] S. Chavan and N. Choubey, “An automated diabetic
Learning Features from Pre-trained Deep Learning Networks,” retinopathy of severity grade classification using transfer
pp. 171–175, 2022. learning and fine-tuning for fundus images,” Multimed. Tools
[9] C. M. Dasari, “A Novel Deep Learning Framework for Appl., 2023, doi: 10.1007/s11042-023-15135-0.
Diabetic Retinopathy Detection,” 2022.
[10] S. S. Mondal, N. Mandal, K. K. Singh, A. Singh, and I. Izonin,
“EDLDR: An Ensemble Deep Learning Technique for
Detection and Classification of Diabetic Retinopathy,”

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on February 12,2025 at 08:17:31 UTC from IEEE Xplore. Restrictions apply.