Detection of Diabetic Retinopathy Using MobileNet Model

U Sadhana

, Tripty Singh

and Beena B. M

Department of Computer Science and Engineering, Amrita School of Computing,

Amrita Vishwa Vidyapeetham, Bengaluru, India

Keywords:

Diabetic Retinopathy, Data Augmentation, MobileNet, IDRiD Dataset.

Abstract:

Diabetic Retinopathy is a chronic disease that may cause blindness to diabetic patients.The proposed system

displays various pathological changes and identiﬁes DR grades for ophthalmologists. The collection of 516

retinal fundus photographs is freely available. We start by removing noise, improving image quality, and stan-

dardizing retinal image sizes. Second, we distinguish between healthy and diabetic retinopathy instances, and

data augmentation is used to increase the volume, quality, and diversity of training data. Next, we divided the

data into three datasets: training, testing, and validation.According to the degree of DR, images are divided

into four groups normal-class 0, mild-class 1, moderate-class 2, and severe-class 3. The proposed method

detects the presence of DR using ﬁne-tuned MobileNet model .This system achieved precision of 91.70%, Re-

call of 89.53%, F-Score of 88.50% and moreover an accuracy of 89.53% for IDRiD dataset. The experiments

yield good results when compared to other systems.

1 INTRODUCTION

A chronic illness that affects millions of individuals

globally is diabetes. Diabetic Retinopathy (DR) is the

term used to describe people whose diabetes causes

an eye condition. The most common cause of vision

loss and blindness in humans is DR.Diabetic retinopa-

thy occurs in two stages: proliferative retinopa-

thy (PDR) and non-proliferative retinopathy (NPDR).

The early stage of diabetic retinopathy is known as

NPDR, and the advanced stage is known as PDR. Five

phases (0–4) can be used to categorize the severity of

DR. There is no retinopathy(0), moderate NPDR(1),

severe NPDR(2), mild NPDR(1), and PDR (4). The

frequency and severity of different related lesion pre-

sentations and outcomes are the primary grading fac-

tors.

Figure 1: Stages of a Diabetic Retinopathy

In the literature on DR diagnosis, we often look

for one or more pre-selected DR-related characteris-

https://orcid.org/0000-0003-4224-0486

https://orcid.org/0000-0002-3688-4392

https://orcid.org/0000-0001-9108-7073

tics in color fundus pictures, such as microaneurysms,

cottonwool spots, hard exudates, and Neo Vascular-

ization. In recent years, computational approaches for

automatically detecting DR severity in fundus images

have been developed. These approaches predict the

existence or absence of DR severity levels, such as

early stage and advanced stage. Approximately one-

third of diabetic people acquire DR, with one-tenth

experiencing severe vision-threatening DR. Initially,

DR patients may not experience any visible symp-

toms. However, when the lesions progress to a more

severe stage, patients may notice eye issues. (Exam-

ples: dark shadow, eyesight loss, etc.) Techniques

must be durable, accurate, and cost-effective. This

approach aims to identify and characterize DR based

on severity. Figure 1 shows the structure of a normal

retina and different stages of DR. Microaneurysms are

little swellings that occur within the blood vessels of

the retina during the ﬁrst stage. Hemorrhage occurs

when blood vessels burst and release blood.Cotton

wool spots (CWS), also known as exudates, occur

during the second stage. These are brought on by

ﬂuid—which includes lipids, white blood cells, ﬁb-

rin, and serum—seeping out of the blood vessels. Ex-

udates are categorized as brilliant lesions because of

their yellowish white tone, while microaneurysms and

hemorrhages are classiﬁed as dark lesions because of

their red appearance. The two lesions differ in size,

478

Sadhana, U., Singh, T. and B. M, B.

Detection of Diabetic Retinopathy Using MobileNet Model.

DOI: 10.5220/0013595100004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 478-483

ISBN: 978-989-758-763-4

shape, and brightness. Because there aren’t enough

ophthalmologists in India, screening each patient by

hand takes time. The indian Diabetic Retinopathy im-

age consists of 512 images. The resolution of this im-

ages are 4288 X 2848 pixels. This dataset composed

of 5 DR and 3 DME. It provides severity level of DR

and DME for each image in the dataset. It also pro-

vides normal retina and DR lesions structures.

2 LITERATURE SURVEY

HUA et al., a proposed design called TFA-Net, which

is a Twofold Feature Augmentation mechanism con-

nected to a backbone convolutional network.Several

convolution blocks are used in the former to extract

representational information at different scales(Bilal

et al., 2021). The latter is built in two stages: ﬁrst,

a Reverse Cross-Attention (RCA) stream is deployed,

and then weight-sharing convolution kernels are em-

ployed.

M. M. Abdelsalam, M. A. Zahran et al., proposed

a detailed explanation of a revolutionary multifrac-

tal geometry-based early DR detection technique is

provided. Image analysis using OCTA (macular

optical coherence tomography angiography) for the

early detection of non-proliferative diabetic retinopa-

thy (NPDR)(Chaudhary and Pachori, 2022).

X. Zeng et al., proposed Automated diagnosis of di-

abetic retinopathy can be achieved by dividing color

retinal fundus photos into two categories(Dharmana

and Aiswarya, 2020).This research describes the use

of transfer learning to train a unique convolutional

neural network model with a Siamese-like topology.

L. Qiao et al., suggested a system that uses convo-

lutional neural network algorithms to analyze fundus

images for the presence of microaneurysms. Deep

learning is incorporated as a key component, and the

system is accelerated by GPUs (Graphic Processing

Units)(Hua et al., 2020).

K. Shankar et al. proposed using a brand-new auto-

mated Hyperparameter Tuning Inception-v4 (HPTI-

v4) model to recognize and categorize DR in color

fundus pictures(Shankar et al., 2020). The contrast

limited adaptive histogram equalization (CLAHE)

model will be used during the preprocessing stage

to raise the fundus image’s contrast level. The pre-

processed image is then segmented using a segmenta-

tion model based on histograms.

J. Wang et al., proposed a retinal fundus image can

be used to directly identify one or more fundus ill-

nesses using a multi-label classiﬁcation ensemble

model based on CNN. Each and every model has

two components. The second section includes a pro-

prietary classiﬁcation neural network for multi-label

classiﬁcation challenges, whereas the ﬁrst uses an

EfﬁcientNet-based feature extraction network. Ulti-

mately, the ﬁnal recognition result is a fusion of the

output probabilities from various models. Addition-

ally, training and testing were conducted using the

data set made available by the Peking University In-

ternational Competition on Ocular Disease Intelligent

Recognition (ODIR 2019)(Wang et al., 2020b).

Juan Wang et al., proposed a hierarchical multi-

task deep learning architecture for diagnosing fun-

dus photos’ DR-related properties and severity con-

currently(Wang et al., 2020a). To account for the

random relationship between DR severity levels and

DR-related features, a hierarchical framework is pro-

posed.

M. D. Alahmadi et al. created a deep neural network

that employs style and content recalibration to scale

informative regions for diabetic retinopathy classiﬁ-

cation.To draw emphasis to texture details in the style

representation, the texture attention module applies a

high-pass ﬁlter. To identify the most informative area

of the input image, the spatial normalization module

uses a convolutional approach(Alahmadi, 2022).

W. Nazih et al.proposed an automated method for de-

termining the severity of DR in fundus images. To

ﬁnd long-range correlations in images, we developed

a vision transformer deep learning pipeline(Nazih

et al., 2023). To train a large vision model on a lim-

ited dataset, the researchers employed transfer learn-

ing. The new real-world FGADR dataset was used to

train the model in order to test it.

ZHOU et al. developed a methodology for gener-

ating high-resolution DR images that performs well

with grading and lesion data. Synthesized data can

improve grading model performance, especially for

photos with high DR levels.(Zhou et al., 2020)

Natarajan Chidambaram et al., focused automated

CAD system that can identify and categorize

exudates in DR. Prior research mostly concen-

trated on using region-based techniques, such as

the Hough transform, watershed transform, re-

gion growth approaches, etc., to segment the optic

disc.(Chidambaram and Vijayan, 2018)

Bindhumol et al. system makes use of Transfer Learn-

ing methods, including ResNet50 and EfﬁcientNetB5.

When comparing the two models classiﬁcation and

confusion matrix results, it was found that ResNet50

performed better at classifying the DR images than

EfﬁcientNetB5.(Bindhumol et al., 2022)

Meher Madhu Dharmanan et al. focused on blob de-

tection and image preprocessing are used to present

an effective, straightforward, and precise feature ex-

traction technique. In the proposed paradigm, testing

Detection of Diabetic Retinopathy Using MobileNet Model

479

is carried out on a scale of 0 to 4; 0 (no DR), 1 (mild

DR), 2 (moderate DR), 3 (severe DR), and 4 (prolifer-

ative DR) to expedite disease identiﬁcation of diabetic

retinopathy.(Dharmana and Aiswarya, 2020)

Praveena S et al. suggested method for diagnosing

diabetic retinopathy will help to produce segmenta-

tion ﬁndings that are accurate with few misclassiﬁ-

cations and will also direct doctors in simulating the

patient’s course of treatment in order to grade the con-

dition. Even though CWS is quicker, better segmenta-

tion outcomes are produced for the LSC approach be-

cause, using a straightforward local feature-based al-

gorithm, it meets both boundary adherence and perse-

verance of global picture structure.(Praveena and La-

vanya, 2019)

Payel Patra et al., proposed a system for diagnosing

DR disease automatically.This study used fundus im-

ages to identify diabetic retinopathy using Resnet50

and Inception V3 architecture. Based on the upcom-

ing output, we were able to achieve an afﬁrmation del-

icacy of 83 percentile.(Patra and Singh, 2022)

L Sai Prajeeth Reddy et al., a variety of algo-

rithms,to extract both basic and sophisticated infor-

mation that aid in the early detection of the condition,

enabling ophthalmologists to diagnose diabetic pa-

tients more accurately and restore their vision.(Reddy

et al., 2022)

3 METHODOLOGY

1. MobileNet Convolutional neural network (CNN)

architecture known as MobileNet was created to

operate effectively on embedded and mobile de-

vices with constrained computational power.

2. MobileNet Architecture MobileNet substitutes

depthwise separable convolutions for traditional

convolutions. In order to achieve this, the standard

convolution is divided into two distinct layers: a

pointwise convolution and a depthwise convolu-

tion. As a result, the network becomes lighter and

uses fewer computations and parameters. Mo-

bileNet Architecture is displayed in Fig.2. De-

ﬁnes the input shape and initializes the MobileNet

base model without the upper classiﬁcation lay-

ers using pre-trained weights from the ”imagenet”

dataset..Constructs custom deep layers for ﬁne-

tuning based on the fundamental MobileNet idea.

This consists of a ﬁnal dense layer with soft-

max activation for classiﬁcation, a dense layer

with ReLU activation, and a global average pool-

ing layer.It describes the outputs and inputs that

go into deﬁning the ﬁnal model. The architec-

ture of the proposed system is divided into four

stages. The initial stage, known as preprocessing,

involves taking out illumination, noise, and arti-

facts from color fundus images. Using the median

ﬁlter, we may enhance the quality of color fundus

images by removing the green channel from an

RGB image, which can then be utilized for binary

classiﬁcation and feature extraction. The second

Figure 2: MobileNet Architecture

phase, feature extraction, is used to retrieve color

fundus images and differentiate between normal

and DR instances. Four DR classes are catego-

rized in the third step, segmentation: mild NPDR,

moderate NPDR, severe NPDR, and PDR. The

fourth stage, known as binary classiﬁcation or

multilabel classiﬁcation, uses the previous feature

vector to categorize the different stages of DR and

to separate the images into normal and DR cases.

3. Preprocessing of image data One of the most im-

portant viewpoints is image preprocessing, which

is vital for altering the data to either ﬁx security

vulnerabilities in the data brought about by the

capturing equipment or to change the

Figure 3: Image Preprocessing Steps

data into a conﬁguration that will be managed

much more successfully and effectively.

4. Morphological Transformation

Morphology is the set of techniques that can be

applied to either the pre-processing of the im-

age segmentation stage’s input data or the post-

processing of its output. To put it another way,

morphological operations can be applied after

segmentation is ﬁnished to eliminate ﬂaws in the

segmented image and provide details about its

composition.

5. Testing and Training

After preprocessing the entire training set, the fea-

tures are taken out of the bleeding areas and ex-

udates to begin the training process. 80% im-

ages for training,20% for testing are used.The tar-

get classes provide all of the training features that

INCOFT 2025 - International Conference on Futuristic Technology

480

Table 1: Summary of Recent Studies on Retinal Disease Detection. (AUC: Area under the curve, ACC: Accuracy, SPE:

Speciﬁcity, SEN: Sensitivity)

Ref. Dataset Images Classes Partitioning Techniques Performance (%)

3 IDRiD 516 DR (0-4), DME (0-3) 80% train, 20% test SVM, KNN, BT Acc: 98.06, Sens: 83.67, Spec: 100

1 DRIVE 40 7 DR cases, 33 healthy 80/20 split ML-CAD System (Deep Learning) Acc: 95.1, AUC: 91.9, Sens: 86.1, Spec: 86.8

8 Shenzhen SiBright 89,917 NPDR, PDR Train: 77,626; Val: 6,200; Test: 4,502 & 1,589 Hierarchical Deep Learning AUC: 95.00

7 ODIR 2019 3,500 patients N, D, G, C, A, H, M, O Uneven EfﬁcientNetB3 Acc: 89, Precision: 63, Recall: 58, AUC: 73, Kappa: 49, F1: 89

6 MESSIDOR 1,200 Normal, Stage 1-3 10-fold CV HPTI-v4 Sens: 98.04, Spec: 99.62, Acc: 99.42, Precision: 97.40

5 KHUMC & Messidor 297 pairs, 1,200 No DR, Mild NPDR, Mod NPDR, Severe NPDR, PDR 1,080 train, 120 test ConvNet + Feature Augmentation Kappa: 90.2, Acc: 94.8, AUC: 99.4

4 IDRiD, Messidor, APTOS 2019 516, 1,200, 3,662 DR (0-4), DME (0-2) 413 train, 103 test 2-D-FBSE-FAWT ACYavg: DR (95.5), DME (96.5)

2 ODIR, SSL, GTest 10,000, 2,023, 506 N, D, G, C, A, H, M 9,000 train/val MCG-Net, MCGS-Net Precision: (64.31, 65.88), Recall: (59.04, 61.60), Kappa: (55.27, 57.65), F1: (89.16, 89.67), AUC: (76.88, 78.16)

Figure 4: A System for Recognizing and Handling DR

Cases

Figure 5: Morphological Transformation Images

were gathered from all of the images to each of

the three classiﬁers independently. The classiﬁers

are lastly kept for testing. Similarly, the prepara-

tion stage also starts the testing process. Further-

more, the prediction values of each classiﬁer are

regarded as votes, and the mode of votes is deter-

mined.

4 EVALUATION METRICS

Classiﬁer Recall Precision

F-

Score

Accuracy

MobileNet 0.895 0.917 0.885 0.895

The performance of a classiﬁer is evaluated

using the Accuracy, Precision, Recall, F1-Score, and

Speciﬁcity.

Accuracy:

The percentage of correctly categorized samples

among all samples is known as accuracy. It is de-

scribed as:

Accuracy =

T P + T N

T P + T N + FP + FN

Precision:

The percentage of true positive predictions among all

positive predictions is measured by precision, also

known as positive predictive value:

Precision =

T P

T P + FP

Recall:

The percentage of true positive predictions among all

real positive samples is measured by recall, which is

often referred to as sensitivity or true positive rate:

Recall =

T P

T P + FN

F1-Score:

The F1-Score provides a single statistic that balances

Precision and Recall by taking the harmonic mean of

the two:

F1-Score = 2 ×

Precision × Recall

Precision + Recall

Speciﬁcity:

Speciﬁcity (also called True Negative Rate) measures

the proportion of true negative predictions out of all

actual negative samples:

Speciﬁcity =

T N

T N + FP

Deﬁnitions:

Detection of Diabetic Retinopathy Using MobileNet Model

481

• The number of accurately predicted positive sam-

ples is known as True Positives (T P).

• The number of accurately predicted negative sam-

ples is known as True Negatives (T N).

• The number of falsely anticipated positive sam-

ples is known as False Positives (FP).

• The amount of negative samples that were mispre-

dicted is known as False Negatives (FN).

5 RESULTS AND DISCUSSION

This paper focuses on CNN model MobileNet, which

had already been trained and reﬁned using the DR

dataset. Fundus images were divided into ﬁve severity

levels by the algorithm, which went from early stage

to advanced stage in DR. CNN automatically extracts

features from retinal images.To improve the network

even more, a fully linked layer is added after the two

layers used for feature extraction and selection. The

suggested method used the pre-trained CNN model

MobileNet,EfﬁcientNetB0 which was ﬁne-tuned with

the DR dataset. Fundus images were divided into ﬁve

severity levels by the algorithm, which went from no

DR to proliferative DR. CNN uses retinal scans to au-

tomatically extract characteristics. After the two lay-

ers used for feature extraction and selection, a fully

linked layer is added to further enhance the network.

First, we load the MobileNet architecture pretrained

on ImageNet, removing its top classiﬁcation layer,

and freezing its weights. This allows us to build a se-

quential model.Then, we add a GlobalAveragePool-

ing2D layer to reduce spatial dimensions and ﬂat-

ten features into a vector. Two Dense layers follow,

each employing ReLU activation and dropout for en-

hanced feature transformation. Finally, a Dense layer

for classiﬁcation is attached with softmax activation.

The Adam optimizer with categorical cross-entropy

loss is used to construct the model. Training is done

with the retrieved features and labels, and variables

like batch size and epochs are adjusted based on the

particular issue.

6 CONCLUSION

The proposed approach uses a scale of 0 to 4 to di-

agnose diabetic retinopathy, In this case, 0 denotes

no DR, 1 mild DR, 2 moderate DR, 3 severe DR,

and 4 proliferative DR. After analyzing various clas-

siﬁcation algorithms, it was discovered that the CNN

model MobileNet which is ﬁne-tuned with DR dataset

is more efﬁcient with an accuracy rate of 89.53%.

Figure 6: Training and Validation Accuracy graph

Figure 7: Training and Validation Loss graph

REFERENCES

Alahmadi, M. D. (2022). Texture attention network for

diabetic retinopathy classiﬁcation. IEEE Access,

10:55522–55532.

Bilal, A., Sun, G., Li, Y., Mazhar, S., and Khan, A. Q.

(2021). Diabetic retinopathy detection and classi-

ﬁcation using mixed models for a disease grading

database. IEEE Access, 9:23544–23553.

Bindhumol, M., Singh, T., and Duraisamy, P. (2022). Di-

abetic retinopathy classiﬁcation using transfer learn-

ing techniques. In 2022 13th International Confer-

ence on Computing Communication and Networking

Technologies (ICCCNT), pages 1–7. IEEE.

Chaudhary, P. K. and Pachori, R. B. (2022). Automatic di-

agnosis of different grades of diabetic retinopathy and

diabetic macular edema using 2-d-fbse-fawt. IEEE

Transactions on Instrumentation and Measurement,

71:1–9.

Chidambaram, N. and Vijayan, D. (2018). Detection of

exudates in diabetic retinopathy. In 2018 Interna-

tional Conference on Advances in Computing, Com-

munications and Informatics (ICACCI), pages 660–

664. IEEE.

Dharmana, M. M. and Aiswarya, M. (2020). Pre-diagnosis

of diabetic retinopathy using blob detection. In 2020

Second International Conference on Inventive Re-

search in Computing Applications (ICIRCA), pages

98–101. IEEE.

Hua, C.-H., Kim, K., Huynh-The, T., You, J. I., Yu, S.-Y.,

Le-Tien, T., Bae, S.-H., and Lee, S. (2020). Convo-

lutional network with twofold feature augmentation

for diabetic retinopathy recognition from multi-modal

INCOFT 2025 - International Conference on Futuristic Technology

482

images. IEEE Journal of Biomedical and Health In-

formatics, 25(7):2686–2697.

Nazih, W., Aseeri, A. O., Atallah, O. Y., and El-Sappagh,

S. (2023). Vision transformer model for predict-

ing the severity of diabetic retinopathy in fundus

photography-based retina images. IEEE Access,

11:117546–117561.

Patra, P. and Singh, T. (2022). Diabetic retinopathy de-

tection using an improved resnet50-inceptionv3 struc-

ture. In 2022 13th International Conference on Com-

puting Communication and Networking Technologies

(ICCCNT), pages 1–6. IEEE.

Praveena, S. and Lavanya, R. (2019). Superpixel based

segmentation for multilesion detection in diabetic

retinopathy. In 2019 3rd International Conference on

Trends in Electronics and Informatics (ICOEI), pages

314–319. IEEE.

Reddy, L. S. P., Ram, K. B. S., Vemu, N. V., Reddy, G. J.,

and Singh, T. (2022). A multifeatured based diabetic

retinopathy detection using hybrid-net. In 2022 13th

International Conference on Computing Communica-

tion and Networking Technologies (ICCCNT), pages

1–6. IEEE.

Shankar, K., Zhang, Y., Liu, Y., Wu, L., and Chen, C.-H.

(2020). Hyperparameter tuning deep learning for di-

abetic retinopathy fundus image classiﬁcation. IEEE

Access, 8:118164–118173.

Wang, J., Bai, Y., and Xia, B. (2020a). Simultane-

ous diagnosis of severity and features of diabetic

retinopathy in fundus photography using deep learn-

ing. IEEE Journal of Biomedical and Health Infor-

matics, 24(12):3397–3407.

Wang, J., Yang, L., Huo, Z., He, W., and Luo, J. (2020b).

Multi-label classiﬁcation of fundus images with efﬁ-

cientnet. IEEE access, 8:212499–212508.

Zhou, Y., Wang, B., He, X., Cui, S., and Shao, L. (2020).

Dr-gan: conditional generative adversarial network

for ﬁne-grained lesion synthesis on diabetic retinopa-

thy images. IEEE journal of biomedical and health

informatics, 26(1):56–66.

Detection of Diabetic Retinopathy Using MobileNet Model

483