Bulletin of Electrical Engineering and Informatics

Vol. 14, No. 2, April 2025, pp. 1063~1072

ISSN: 2302-9285, DOI: 10.11591/eei.v14i2.8248  1063

Efficient diabetic retinopathy detection using deep learning

approaches and Raspberry Pi 4

Silpa Ajith Kumar1, James Satheesh Kumar1, Sharath Chandra Bharadwaj Mahabaleswara2
1
Department of Electronics and Instrumentation Engineering, Dayananda Sagar College of Engineering, Bengaluru, India
2
Medical Director, Atharva Eye Hospital, Bengaluru, India

Article Info ABSTRACT

Article history: Diabetic retinopathy (DR) is a leading cause of vision loss, predominantly
affecting individuals aged 25-74 with diabetes mellitus. Timely medical
Received Jan 29, 2024 intervention can protect against irreversible blindness in over 90% of cases,
Revised Oct 8, 2024 emphasizing effectively identifying and treating DR. In the scope of deep
Accepted Nov 19, 2024 learning (DL), the possibility of using them in DR screening has garnered a
lot of interest. Specifically, we adopted the densely connected convolutional
networks (DenseNet) model because to its capacity to acquire complex
Keywords: features and learn from diverse datasets. Developing the computational
model on retinal images labelled with varying phases of DR are obtained
Convolutional neural networks from databases such as Messidor and Kaggle. To enhance accessibility and
Deep learning user-friendliness, we integrated the DenseNet model into a Raspberry Pi 4, a
DenseNet architecture compact, affordable and widely accessible computing platform. The
Diabetic retinopathy proposed approach resulted in an impressive classification accuracy of 88%,
Fundus images demonstrating its proficiency in distinguishing between different phases of
Raspberry Pi 4 DR progression. The study aims to assist in the early detection and diagnosis
of the disease, providing a potential resource that could help medical
practitioners and ophthalmologists to evaluate the extent of DR in a timely
manner.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Silpa Ajith Kumar
Department of Electronics and Instrumentation Engineering, Dayananda Sagar College of Engineering
Bengaluru, Karnataka, India
Email: silpaajithkumar@gmail.com

1. INTRODUCTION
Diabetic retinopathy (DR) is a prevalent and severe health condition arising from diabetes mellitus.
It impacts the retina by damaging blood vessels, resulting in issues like microaneurysms, hemorrhages, soft,
and hard exudates [1]. Loss of vision usually results from central retinal swelling, leading to impaired vision.
Additionally, the growth of retinal abnormal blood vessel, potentially causing bleeding and retinal scarring,
ultimately resulting in blindness. Swelling in the retina’s central region can disrupt visual function and cause
vision loss. Retinal haemorrhage or scarring from the growth of abnormal blood vessels is another common
cause of blindness [2]. DR stands as a significant and prevalent concern in the realm of public health. This
ocular complication of diabetes mellitus poses a substantial threat to vision, affecting millions of individuals
worldwide. DR manifests many levels, every level with unique characteristics and implications for patient
care. The five stages of DR ranging from mild nonproliferative diabetic retinopathy (NPDR) to severe
proliferative diabetic retinopathy (PDR) represent a continuum of disease progression [3], where early
detection and precise staging hold the key to effective treatment and vision preservation. In addition to
diabetes, high blood pressure and high cholesterol both enhance the likelihood that you may develop DR.
Taking care of your heart and lowering your cholesterol levels will help protect your eyesight [4], [5]. The

Journal homepage: http://beei.org

1064  ISSN: 2302-9285

automatic detection and staging of DR have long been a subject of intense research and development.
Traditional methodologies relied on manual assessments by ophthalmologists, a time-consuming and
resource-intensive process. Nevertheless, convolutional neural networks (CNNs) and other deep learning
(DL) innovations, have revolutionized the field. These cutting-edge techniques utilize the potential of
machine learning to analyze retinal images, offering the promise of efficient, accurate, and scalable DR
diagnosis [6]-[9]. The purpose behind this project set out to uncover thorough investigation the automatic
detection and staging of DR using state-of-the-art DL and CNN techniques. By delving into the existing
literature [10]-[12], we navigate the dynamic environment of DR diagnosis, highlights on the transformative
potential of artificial intelligence (AI-driven) solutions. By including the features of both the entropy of
grayscale images and the unsharp masking for the green channel extraction, the DL approaches used to
fundus images from “Kaggle Diabetic Retinopathy” [13] dataset, are implemented using a bichannel CNN
architecture, thereby boosting the performance measures [14]. AI techniques have applications beyond
ophthalmology in many other medical fields [15], [16]. Persons who have difficulty with urinary tract issues
can now be monitored using a system designed by Kim et al. [17] to assist with risk assessment and
management in a related study, Eun et al. [18] suggests ResNet-50 as the CNN architecture for detecting
urolithiasis, or kidney stones, in ureters.
Many of the researchers failed to take the five DR stages into account when they split DR images
and do the diagnosis only into two categories: no DR and DR in most of the existing works [19], [20].
Clinical specialists in DR rated the AI model that Bajwa et al. [2] used to categorize test images as either
DR-positive or DR-negative with an accuracy of 93.72%, sensitivity of 97.30%, and specificity of 92.90%.
To treat the retina with the appropriate process and prevent deterioration and blindness, it is vital to know the
specific stage of DR, and the DR phases help with that. Monteiro [21] introduced a blended grading predictor
employing ten distinct DL models. Through a 5-fold cross-validation approach, individual models were
trained and their predictions were combined to generate a final score. This strategy aims in reducing the
generalization error inherent in a single DL model, presenting a promising technique for leveraging
information from multiple models.
This review in addition to integrating the recent advancements but also identifies gaps and
challenges that demand further investigation. With this work, we intend to assist in the ongoing attempt for
more effective detection and treatment of DR in its early stages, ultimately preserving the precious gift of
sight for countless individuals. The following is the outline of the paper: in the second section, the materials
and procedures which support the suggested model are discussed. Section 3 depicts the results and the
discussions of the suggested work followed by the conclusion in section 4.

2. METHODS
Developing a DL algorithm that can reliably identify DR in its early stages is the primary aim of the
study that is presented. By streamlining the clinical diagnosis and identification of DR, can be accessible for
clinicians to make decisions quicker.

2.1. Datasets and data preparation

Utilising of multiple datasets has led to advancements in DR detection. These datasets include well-
known ones such as DRIVE, Messidor, Kaggle diabetic retinopathy detection (Kaggle-DRD), e-Ophtha,
APTOS 2019 blindness detection, IDRiD, and Messidor-2 [22]-[24]. Researchers and developers using ML
and DL algorithms have relied on these datasets, which provide a wide range of images representing different
degrees of illnessintensity. The database are incredibly helpful for training and assessing models, which
allows for advancements in accurate and reliable detection methods. Therefore, to guarantee that their models
obtain consistent information, researchers frequently conduct common preprocessing activities including
cropping, resizing, and normalisation. Our research adds to the expanding field of research on DR
identification using the densely connected convolutional networks (DenseNet) model, with a focus on the
Kaggle and Messidor datasets. For DR classification, we can use Kaggle, a large public dataset with over
35,000 high-resolution fundus images that capture DR detection on the retinal cells. These images include
artefacts, blurring, focusing, and exposure difficulties, across a range of devices used at different primary
care facilities. Our goal in introducing this diversity is to make it more realistic. Also, these databases ground
truth labels aren’t completely noise-free, which adds to their authenticity. In line with the ICDRDSS, which
is an international standard for assessing the severity of DR [25], expert graders assessed the images. With its
large size and real heterogeneity, the dataset has enormous potential for building ophthalmology-specific
classification models. Three eye care facilities in France collected 120 images of the retinal fundus between
2005 and 2006 to form the MESSIDOR dataset. There are 1,058 original photos within the following
Messidor 2 collection. These datasets do not include pixel-wise lesion segmentation data, but they do provide

Bulletin of Electr Eng & Inf, Vol. 14, No. 2, April 2025: 1063-1072
Bulletin of Electr Eng & Inf ISSN: 2302-9285  1065

imaging-based medical diagnosis, particularly about DR severity. Also, unlike the popular ICDRSS
procedure, this dataset uses its own method of evaluation [26].

2.2. Pre-processing steps

Pre-processing is a crucial first stage of utilizing DL approaches to DR detection, with the goal of
improving data quality and consistency to prepare it for model training [27]. Several common pre-processing
steps are typically undertaken are image acquisition, image annotation, image enhancement, image cropping
and resizing, data augmentation, and data pre-processing respectively. With our work, we made use of the
Messidor and open-source databases for DR identification. These datasets included annotated retinal images
indicating various phases of DR. Pre-processing steps involved cropping, resizing, and unsharp masking of
image data. Resizing helps improve the speed which involves effectiveness of the model by cutting down on
data volume, making especially helpful for big datasets or resource-constrained environments. It ensures
consistent and accurate image processing, a requirement for many DL methods. The Messidor dataset was
resized to 880×850 pixels, the Messidor 2 dataset to 674×680 pixels and Kaggle to 1,024×1,024 pixels of
resolution. Image cropping aims to improve model accuracy and efficiency in order to decrease background
noise, remove unnecessary sections and facilitating the model concentrates on crucial features. This step
frequently results in improved model performance and decreased false positives or negatives. Unsharp
masking, while not commonly used DR detection with DL, can sharpen the image and its clarity. It involves
applying a sharpening filter to accentuate edges and fine details, potentially improving image quality. If
considered as collectively, these pre-processing procedures facilitate DL model’s overall effectiveness and
reliability in DR disgnosis. The Figure 1 depicts the process flow of the DR detection, extraction of features
and assessments of the performance measures.

Figure 1. Procedure for DR identification and feature extraction

2.3. Contrast limited adaptive histogram equalization algorithm

For enhancement of images, one method is to use contrast limited adaptive histogram equalization
(CLAHE) [28] that aims to enhance clarity. It operates by expanding the image’s contrast range, aligning the
lightest and darkest pixels with the lowest and highest potential values. In order to accomplish this, CLAHE
first partitions the image into smaller areas, known as “tiles”, along with histogram equalization applied to
every single tile independently. The approach preserves local contrast within each tile while improving the
total contrast of the image. In our study, we applied CLAHE to green-featured images to enhance their
quality. As illustrated in Figure 2, the fundus pictures had their red, green, and blue channels removed. With
the colour data separated, the model can access the finer details in each channel, which could help it better
identify important traits and patterns associated with the various phases of DR. Before the data is subjected to
additional analysis and model training, this stage helps enhance and refine it.

Efficient diabetic retinopathy detection using deep learning approaches and … (Silpa Ajith Kumar)
1066  ISSN: 2302-9285

Figure 2. Extraction the fundus images of red, green, and blue channels

Since the Green channel was outperforming the other two, we decided to apply CLAHE on its
featured images in order to improve their quality. The pre-processed images following the implementation of
CLAHE algorithm can be viewed in Figure 3 which succinctly explains the visual output obtained after
applying CLAHE to retinal images within the context of DR analysis. The Figure 3 showcases the enhanced
visual representation achieved through the advanced image processing technique of CLAHE.

Figure 3. Image preprocessing using the CLAHE method on the green channel

2.4. Proposed algorithm and architecture of the convolutional neural network model
To address issues with deep networks, an architecture for CNNs called DenseNet was developed. By
adding dense connections between layers, it improves information flow and solves problems like the
vanishing gradient problem. Figure 4 depicts the DenseNet design, which has a dense connectivity structure
between layers to encourage feature reuse and gradient flow network-wide.

Figure 4. DenseNet architecture

Bulletin of Electr Eng & Inf, Vol. 14, No. 2, April 2025: 1063-1072
Bulletin of Electr Eng & Inf ISSN: 2302-9285  1067

Typically, in traditional CNNs, the feature maps created by one layer are used as input by the next
layer [29]. However, DenseNet employs a novel approach to connection establishment, with each layer being
feed-forward associated with each following layer. As a consequence, there are many interconnected layers,
with each subsequent layer receiving its output as an input. During backpropagation, this method guarantees
that all layers directly get gradients from all subsequent layers. This function encourages better gradient flow
throughout the system and helps with the vanishing gradient problem.
A number of dense blocks, each with a number of layers, make up DenseNet. Each layer in a dense
block forms a subnetwork that is highly connected to all the other levels in that block. By improving the
network’s information flow and allowing for efficient feature reuse, these dense links may increase
representational capacity of the network as a whole.

2.5. Training the DenseNet model

Stochastic gradient descent (SGD) and adaptive optimizers like Adam are used to train DenseNet,
which focuses on optimization and backpropagation. When solving a classification problem, for example,
cross-entropy loss would be the appropriate choice of loss function.
a. Dense connectivity
Let “𝐻𝑙 ” denote the output layer “l”, and the input to layer “l”, is the concatenation of the feature
maps from all preceding layers represented as (1):

𝐻𝑙 = [ 𝐻0 , 𝐻1 , … . . , 𝐻𝑙−1 ] (1)

b. Composite function in dense blocks

The output of each layer with a dense block is computed as (2):

𝐻𝑙 = 𝑓𝑙 ([𝐻0 , 𝐻1 , … … . , 𝐻𝑙−1 ] ) (2)

c. Transition layer
The output of the transition layer is calculated as (3):

𝐻𝑙 = 𝐵𝑁 (𝑊𝑙 ∗ 𝐻𝑙−1 ) (3)

where “*” represents convolution, BN represents the batch normalization, and Wl represents the 1×1
convolutional layer.
d. Global average pooling
The global average pooling operation is applied as (4):
1
𝑦𝑘 = ∑𝐻 𝑊 𝑘
𝑖=1 ∑𝑗=1 𝐻 𝑖𝑗 (4)
𝐻𝑥𝑊

where “H” and “W” are height and width of the feature maps, and 𝐻 𝑘𝑖𝑗 is the k-th channel of the feature map
at position (i, j).

3. RESULTS AND DISCUSSION

To train the DenseNet model for the DR, a labelled dataset is essential. This dataset must contain
retinal images accurately labelled with corresponding DR severity levels. Initially, the dataset undergoes
preprocessing to ensure its suitability for training. This typically involves resizing images to a consistent
resolution, normalizing pixel values, and potentially augmenting the data to enhance diversity and the
model’s generalization. Augmentation techniques like rotation, flipping, and zooming are applied to create
various image variations. Configuring the DenseNet model architecture follows. This entails specifying the
number of layers, growth rate, and other hyperparameters defining the network’s structure. Selecting
hyperparameters often involves experimentation and fine-tuning for optimal performance. During the training
process, the labeled dataset is composed into mini-batches, each fed into the DenseNet model. Model output
is compared to ground truth labels using a suitable loss function, with categorical cross-entropy being a
common choice for DR classification as depicted in Figure 5.
This loss quantifies dissimilarity between predicted and true label distributions. During training, the
DenseNet model iteratively updates parameters by backpropagating gradients through the network and
adjusting weights determined by the chosen optimization algorithm This continues on for a number of epochs,
with a pass occurring at the end of each epoch through the entire training dataset. In most cases [30], [31],

Efficient diabetic retinopathy detection using deep learning approaches and … (Silpa Ajith Kumar)
1068  ISSN: 2302-9285

tracking the model’s performance dictates the number of epochs on a separate validation set, stopping when
performance reaches a satisfactory level or starts to deteriorate.

Figure 5. Categorical cross entropy loss

Categorical accuracy is another performance measure which is a quantitative measure of how many
samples in a dataset have been correctly categorized. When working with various classes, decides the overall
accuracy of classification. A better-performing model will have higher category accuracy. Precision is
another metric that quantifies the proportion of true positive predictions (correctly predicted positive
samples) out of the total predicted positive samples. It measures the model’s ability to avoid false positives.
Precision is calculated as true positives divided by the sum of true positives and false positives . This is
depicted in Figure 6 as the Categorical performance measures of accuracy and precision. The Figure 6(a)
represents the categorical accuracy. The performance measure in terms of precision for the proposed model is
shown in Figure 6(b).

(a)

(b)

Figure 6. Categorical performance measures of the proposed model: (a) accuracy and (b) precision

Bulletin of Electr Eng & Inf, Vol. 14, No. 2, April 2025: 1063-1072
Bulletin of Electr Eng & Inf ISSN: 2302-9285  1069

A potential method for assessing how well a classification model is performing is by examining at
its confusion matrix. The information illustrates the total number of accurate predictions for each class,
including negative as well as positive results. Figure 7 represents the confusion matrix and the performance
measures of the proposed approach using DenseNet algorithm. The Figure 7(a) depicts the confusion matrix
for the Kaggle and Messidor datasets. The performance measures of the algorithm are represented in
Figure 7(b) in terms of accuracy, precision, recall, and F1 score.

(a)

(b)

Figure 7. Proposed results of DenseNet architecture: (a) confusion matrix and (b) performance measures

3.1. Integration with Raspberry Pi 4

Controlling LEDs to indicate DR stage is made possible through the integration of the proposed
work with a Raspberry Pi. As part of this integration, the classification system is connected to the Raspberry
Pi, which can then receive information from the system and turn on LEDs that correlate to different levels of
disease severity. The first step is to get the Raspberry Pi ready with all the necessary software and libraries.
Configuring the Pi for development and installing an operating system (such as Raspbian) are part of this
process. Since the Raspberry Pi’s GPIO pins are what really connect to the LEDs, it’s also necessary to
install libraries that allow for control of these pins. After that, you should update the Raspberry Pi 4 with the
DenseNet model that has been trained. Making a connection between the categorization system and the Pi
allows this to happen. After receiving stage predictions, the Raspberry Pi can control the LEDs through GPIO
pins. Users can control the status of each DR stage-corresponding LED by setting its corresponding GPIO
pin to high (on) or low (off). Figure 8 demonstrates how to efficiently manage the GPIO pins and LED
conditions using the tools provided in programming languages like Python.

Efficient diabetic retinopathy detection using deep learning approaches and … (Silpa Ajith Kumar)
1070  ISSN: 2302-9285

Figure 8. Integration with Raspberry Pi 4

4. CONCLUSION
DL methods exhibit significant potential in diagnosing and detecting early-stage DR compared to
traditional methods. DL’s performance improves as the volume of available databases increases, offering
substantial opportunities for physicians to analyze, screen, and draw insights from retina image
datasets.While DL techniques have made notable progress in DR classification, there’s room for enhancing
performance metrics. This can be achieved by incorporating real-time clinical datasets and integrating
hardware modules with DL techniques. Improving DR classification models with cutting-edge approaches
and DL should be the goal of ongoing research. With the help of modern technology, patients and healthcare
providers can connect remotely, opening up new possibilities such as recording retinal images at home and
conducting assessments in real-time using data collected in real-time.
Researchers can gain insight into DR and develop more targeted treatments by combining Raspberry
Pi retinal images with patient data such as medical history and genetic information. Predictive models for
customised preventive and management suggestions can be developed by broadening the scope of the current
research to include other risk factors, such as blood glucose levels and the duration of diabetes. Better patient
outcomes and less condition-related vision loss are on the horizon attributed to these enhancements. The goal
of this objectives is to help with early disease detection and management by giving ophthalmologists and
other healthcare providers an efficient method for quickly and accurately assessing the severity of DR. This
has the potential to help improve patient outcomes and prompt interventions to avoid diabetic eye disease.

ACKNOWLEDGEMENTS
This study wouldn’t have been possible without the resources and facilities offered by the Research
Centre of Electronics and Instrumentation Engineering department of Dayananda Sagar College of
Engineering and Atharva Eye Hospital [Ref No.AEH/23-24/011] for grading the fundus images, which the
authors are deeply grateful to. Also, we’d like to express our gratitude to the journal’s reviewers, whose
insightful comments greatly enhanced the quality of this work.

REFERENCES
[1] N. Tsiknakis et al., “Deep learning for diabetic retinopathy detection and classification based on fundus images: A review,”
Computers in Biology and Medicine, vol. 135, pp. 1–19, 2021, doi: 10.1016/j.compbiomed.2021.104599.
[2] A. Bajwa, N. Nosheen, K. I. Talpur, and S. Akram, “A Prospective Study on Diabetic Retinopathy Detection Based on Modify
Convolutional Neural Network Using Fundus Images at Sindh Institute of Ophthalmology & Visual Sciences,” Diagnostics, vol.
13, no. 3, pp. 1–11, Jan. 2023, doi: 10.3390/diagnostics13030393.
[3] S. D. Solomon and M. F. Goldberg, “ETDRS Grading of Diabetic Retinopathy: Still the Gold Standard?,” Ophthalmic Research,
vol. 62, no. 4, pp. 190–195, 2019, doi: 10.1159/000501372.
[4] A. Rodtook and S. Chucherd, “Optic disc localization using graph traversal algorithm along blood vessel in polar retinal image,”
Bulletin of Electrical Engineering and Informatics, vol. 11, no. 6, pp. 3301–3312, Dec. 2022, doi: 10.11591/eei.v11i6.3618.
[5] K. Kangra and J. Singh, “Comparative analysis of predictive machine learning algorithms for diabetes mellitus,” Bulletin of
Electrical Engineering and Informatics, vol. 12, no. 3, pp. 1728–1737, Jun. 2023, doi: 10.11591/eei.v12i3.4412.
[6] Y. S. Devi and S. P. Kumar, “A deep transfer learning approach for identification of diabetic retinopathy using data

Bulletin of Electr Eng & Inf, Vol. 14, No. 2, April 2025: 1063-1072
Bulletin of Electr Eng & Inf ISSN: 2302-9285  1071

augmentation,” IAES International Journal of Artificial Intelligence (IJ-AI), vol. 11, no. 4, pp. 1287–1296, Dec. 2022, doi:
10.11591/ijai.v11.i4.pp1287-1296.
[7] A. M. Mutawa, S. Alnajdi, and S. Sruthi, “Transfer Learning for Diabetic Retinopathy Detection: A Study of Dataset
Combination and Model Performance,” Applied Sciences (Switzerland), vol. 13, no. 9, pp. 1–17, May 2023, doi:
10.3390/app13095685.
[8] D. Wang and L. Wang, “On OCT image classification via deep learning,” IEEE Photonics Journal, vol. 11, no. 5, pp. 1–14, Oct.
2019, doi: 10.1109/JPHOT.2019.2934484.
[9] M. Al-Smadi, M. Hammad, Q. B. Baker, and S. A. Al-Zboon, “A transfer learning with deep neural network approach for diabetic
retinopathy classification,” International Journal of Electrical and Computer Engineering, vol. 11, no. 4, pp. 3492–3501, 2021,
doi: 10.11591/ijece.v11i4.pp3492-3501.
[10] A. Pradeep and X. F. Joseph, “Binary operation based hard exudate detection and fuzzy based classification in diabetic retinal
fundus images for real time diagnosis applications,” International Journal of Electrical and Computer Engineering, vol. 10, no. 3,
pp. 2305–2312, 2020, doi: 10.11591/ijece.v10i3.pp2305-2312.
[11] P. Datta, P. Das, and A. Kumar, “Hyper parameter tuning based gradient boosting algorithm for detection of diabetic retinopathy:
an analytical review,” Bulletin of Electrical Engineering and Informatics, vol. 11, no. 2, pp. 814–824, Apr. 2022, doi:
10.11591/eei.v11i2.3559.
[12] R. C. Wihandika, P. P. Adikara, S. Adinugroho, Y. A. Sari, and F. Utaminingrum, “Retinal blood vessel segmentation using
multiple line operator-based methods,” Bulletin of Electrical Engineering and Informatics, vol. 11, no. 3, pp. 1696–1705, Jun.
2022, doi: 10.11591/eei.v11i3.3026.
[13] G. Quellec, K. Charrière, Y. Boudi, B. Cochener, and M. Lamard, “Deep image mining for diabetic retinopathy screening,”
Medical Image Analysis, vol. 39, pp. 178–193, 2017, doi: 10.1016/j.media.2017.04.012.
[14] S. I. Pao, H. Z. Lin, K. H. Chien, M. C. Tai, J. T. Chen, and G. M. Lin, “Detection of Diabetic Retinopathy Using Bichannel
Convolutional Neural Network,” Journal of Ophthalmology, pp. 1–7, Jun. 2020, doi: 10.1155/2020/9139713.
[15] Y. Cho, J. M. Park, and S. Youn, “General Overview of Artificial Intelligence for Interstitial Cystitis in Urology,” International
Neurourology Journal, vol. 27, pp. S64–S72, Nov. 2023, doi: 10.5213/inj.2346294.147.
[16] M. S. Choo, H. Y. Ryu, and S. Lee, “Development of an Automatic Interpretation Algorithm for Uroflowmetry Results:
Application of Artificial Intelligence,” International Neurourology Journal, vol. 26, no. 1, pp. 69–77, Mar. 2022, doi:
10.5213/inj.2244052.026.
[17] E. S. Kim, D. J. Shin, S. T. Cho, and K. J. Chung, “Artificial Intelligence-Based Speech Analysis System for Medical Support,”
International Neurourology Journal, vol. 27, no. 2, pp. 99–105, Jun. 2023, doi: 10.5213/inj.2346136.068.
[18] S. J. Eun, M. S. Yun, T. K. Whangbo, and K. H. Kim, “A Study on the Optimal Artificial Intelligence Model for Determination of
Urolithiasis,” International Neurourology Journal, vol. 26, no. 3, pp. 210–218, Sep. 2022, doi: 10.5213/INJ.2244202.101.
[19] W. L. Alyoubi, W. M. Shalash, and M. F. Abulkhair, “Diabetic retinopathy detection through deep learning techniques: A
review,” Informatics in Medicine Unlocked, vol. 20, pp. 1–11, 2020, doi: 10.1016/j.imu.2020.100377.
[20] S. Sengupta, A. Singh, H. A. Leopold, T. Gulati, and V. Lakshminarayanan, “Ophthalmic diagnosis using deep learning with
fundus images – A critical review,” Artificial Intelligence in Medicine, vol. 102, pp. 1–37, Jan. 2020, doi:
10.1016/j.artmed.2019.101758.
[21] F. C. Monteiro, “Diabetic Retinopathy Grading using Blended Deep Learning,” Procedia Computer Science, vol. 219, pp. 1097–
1104, 2023, doi: 10.1016/j.procs.2023.01.389.
[22] M. M. Fraz, M. Badar, A. W. Malik, and S. A. Barman, “Computational Methods for Exudates Detection and Macular Edema
Estimation in Retinal Images: A Survey,” Archives of Computational Methods in Engineering, vol. 26, no. 4, pp. 1193–1220, Sep.
2019, doi: 10.1007/s11831-018-9281-4.
[23] V. Lakshminarayanan, H. Kheradfallah, A. Sarkar, and J. J. Balaji, “Automated detection and diagnosis of diabetic retinopathy: A
comprehensive survey,” Journal of Imaging, vol. 7, no. 9, pp. 1–26, Aug. 2021, doi: 10.3390/jimaging7090165.
[24] C. Rodriguez-Leon, W. Arevalo, O. Banos, and C. Villalonga, “Deep Learning for Diabetic Retinopathy Prediction,” in Lecture
Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol.
12861 LNCS, 2021, pp. 537–546, doi: 10.1007/978-3-030-85030-2_44.
[25] D. S. W. Ting et al., “Deep learning in ophthalmology: The technical and clinical considerations,” Progress in Retinal and Eye
Research, vol. 72, pp. 1–24, Sep. 2019, doi: 10.1016/j.preteyeres.2019.04.003.
[26] Early Treatment Diabetic Retinopathy Study Research Group, “Early Photocoagulation for Diabetic Retinopathy: ETDRS Report
Number 9,” Ophthalmology, vol. 98, no. 5, pp. 766–785, May 1991, doi: 10.1016/S0161-6420(13)38011-7.
[27] S. Dash, S. Verma, Kavita, S. Kim, J. Shafi, and M. F. Ijaz, “Analysis of Vessel Segmentation Based on Various Enhancement
Techniques for Improvement of Vessel Intensity Profile,” Computational Intelligence and Neuroscience, pp. 1–20, Jun. 2022, doi:
10.1155/2022/7086632.
[28] M. Nahiduzzaman et al., “Diabetic retinopathy identification using parallel convolutional neural network based feature extractor
and ELM classifier,” Expert Systems with Applications, vol. 217, pp. 1–11, May 2023, doi: 10.1016/j.eswa.2023.119557.
[29] T. Mahmoudi and A. Mehdizadeh, “Artificial Intelligence in Medicine,” Journal of Biomedical Physics and Engineering, vol. 12,
no. 6, pp. 549–550, Dec. 2022, doi: 10.31661/jbpe.v0i0.2211-1566.
[30] S. A. Kumar and J. S. Kumar, “Retinal lesions classification for diabetic retinopathy using custom ResNet-based classifier,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 33, no. 1, pp. 405–415, 2024, doi:
10.11591/ijeecs.v33.i1.pp405-415.
[31] S. A. Kumar and J. S. Kumar, “A review on recent developments for the retinal vessel segmentation methodologies and exudate
detection in fundus images using deep learning algorithms,” in Advances in Intelligent Systems and Computing, vol. 1108 AISC,
2020, pp. 1363–1370, doi: 10.1007/978-3-030-37218-7_143.

Efficient diabetic retinopathy detection using deep learning approaches and … (Silpa Ajith Kumar)
1072  ISSN: 2302-9285

BIOGRAPHIES OF AUTHORS

Silpa Ajith Kumar is an Assistant Professor in the Department of Electronics

and Instrumentation Engineering, Dayananda Sagar College of Engineering, Bengaluru,
Karnataka, India. She holds an M.E degree in Applied Electronics. Her research areas are
machine learning, deep learning, medical image analysis, biomedical instrumentation, and
automation in process control. She is pursuing a Ph.D. in the area of Biomedical Image
Processing. She has authored or co-authored over 11 internationally refereed articles. Her
research interests include computer vision, artificial intelligence, biomedical instrumentation,
image processing, IoT, medical image analysis, and VLSI design. She can be contacted at
email: silpaajithkumar@gmail.com.

Dr. James Satheesh Kumar received his B.E., from Madurai Kamaraj
University, India, M.Tech., from NIT Trichy, India and Ph.D. degree from Karunya Institute
of Technology and Sciences (Deemed University) India, in 1999, 2005, and 2014,
respectively. He started his career as a teaching faculty during the year 2001. He is currently
associated with Dayananda Sagar College of Engineering, Bengaluru, India as an Associate
Professor. He has authored or co-authored over 25 internationally refereed articles. His current
research interests include biomedical imaging, control engineering, or process control.
Presently he is guiding research scholars from Visvesvaraya Technological University in
interdisciplinary domain. He can be contacted at email: jsatheeshngl@gmail.com.

Dr. Sharath Chandra Bharadwaj Mahabaleswara completed his MBBS from

JJM Medical College, Davangere in 2005 and post graduation in Ophthalmology from Sankara
Eye Hospital, Coimbatore, affiliated to National Board of Examinations, New Delhi in 2009.
He did his long-term clinical fellowship in Glaucoma from Sankara Nethralaya, Chennai in
2011 and PG diploma in medical law and ethics (PGDMLE) from National Law School India
University, Bengaluru. His areas of expertise include cataract surgery, glaucoma and
community ophthalmology. He has attended various State and National conference as faculty
and presented research papers and was actively involved in PG teaching. Currently he is
practicing at Atharva Eye Hospital, Bengaluru as a Senior consultant ophthalmologist. He can
be contacted at email: atharvaeyehospital@gmail.com.

Bulletin of Electr Eng & Inf, Vol. 14, No. 2, April 2025: 1063-1072