AI-Based Early Detection of Diabetic

Retinopathy Using Retinal Images

Anugraha Abigail Flora P(a), Prof. Joseph Jeyaranjan N (b),
(a)(b)
Division of Data Science and Cyber Security, Karunya Institute of Technology and Sciences, Coimbatore, India
Emails: anugrahaabigail@[Link], @[Link]

Abstract— Diabetic Retinopathy (DR) has been recognized as

one of the most common causes of preventable blindness in B. Problem Statement
diabetic patients. Though early diagnosis is crucial in minimizing While many AI-based DR detection models have reported
the risks of vision loss, manual fundus image screening is a time- high accuracy, most act like "black boxes," where
consuming process and highly subjective. Herein, this work puts predictions are made without any visual justification. This
forward an automated deep learning approach for early DR
has raised several concerns regarding clinical trust,
detection using retinal images, while incorporating
EfficientNetB0 for classification and Grad-CAM for visual interpretability, and regulatory approval. Many models even
explainability. In this approach, the model will be trained on face difficulty in early-stage DR detection, where the lesions
preprocessed fundus images and tested on different stages of DR in the retina can be very few or subtle. Thus, there is a need
severity. Grad-CAM visualization generates localized heatmaps for a robust interpretable deep learning model capable of
that highlight pathological regions, thereby enabling transparent
detecting early DR by providing visual evidence of its
and interpretable decision-making. The proposed model shows
excellent performance in classification tasks with clear decision process.
interpretability to bridge the gap between AI predictions and C. Significance of the Study
clinical validation. This can potentially facilitate The proposed work helps narrow the gap between
ophthalmologists in early diagnosis and reduce the workload and automated DR detection and clinical applicability. By
help in large-scale community screening for diabetic retinopathy.
integrating EfficientNetB0 for efficient feature extraction
Index Terms— Diabetic Retinopathy, Deep Learning,
EfficientNetB0, Grad-CAM, Explainable AI, Retinal Image with Grad-CAM for model interpretability, this study
Classification, Medical Imaging, Early Detection enhances both accuracy and transparency in DR diagnosis.
The explainable heatmap visualization will help
I. INTRODUCTION ophthalmologists understand the decision rationale behind
A. Background AI-assisted screening systems, thus building trust in such
Diabetic Retinopathy (DR) is a complication of systems. Additionally, such an approach might enable the
diabetes causing damage to the small blood vessels in the early detection of DR and thus support community-level
retina, which might affect vision. According to the World screening programs, lessen the diagnostic workload, and
Health Organization, approximately one-third of people reduce vision loss among diabetic patients. The proposed
with diabetes develop some stage of DR during their methodology is in line with the wider goal of developing
lifetime. Timely detection and early intervention are interpretable AI systems for medical imaging and public
important to prevent irreversible loss of vision. However, in D. Paper Overview
many regions, the availability of trained ophthalmologists Related work on diabetic retinopathy detection
and facilities for regular screening are still limited. and explainable AI techniques has been reviewed
Recent advances in deep learning have transformed medical in Section II. Section III describes the proposed
image analysis and enabled automated systems to assist methodology, which discusses dataset details,
clinicians in the detection of subtle abnormalities in the preprocessing techniques, model architecture, and
retina. In particular, CNNs have shown remarkable
Grad-CAM-based explainability. Section IV
performance in classifying fundus images and detecting
DR-related lesions. However, their adoption in clinical discusses the experimental setup, results, and
workflow is restricted because models lack transparency visualizations obtained through Grad-CAM.
and it is hard to interpret the prediction. Comparisons with existing models have also been
performed. Section V concludes the study with
key findings, limitations, and potential future
enhancements for real-world deployment.
 II. RELATED WORKS
The identification of DR using deep learning has been studied a 0 – No DR
lot in recent times because of the growing number of large-scale 1 - Mild DR
retinal image datasets and advancements in convolutional neural
2 - Moderate DR
networks. Most early works focused on handcrafted feature
extraction, which suffers from a limited ability to represent 3-Severe DR
complex retinal patterns. With the introduction of deep learning 4 - Proliferative DR
methods, CNN-based architectures have increasingly performed
The dataset has both left and right eye images of diabetic
well for medical image classification tasks.
patients captured under different illumination and focus
One of the first large-scale DR detection systems was developed conditions. The data was stratified to ensure that
by Gulshan et al. [1], where a deep CNN was trained on fundus representation across all classes is balanced; augmentation
photographs. They reported performance comparable to techniques were thus used to minimize class imbalance. The
ophthalmologists. However, their model was not interpretable, images were resized to a size of 224×224 pixels, as the input
which often obscures decision mechanisms from clinicians. Pratt required for EfficientNetB0.
et al. [2] used a CNN architecture to classify DR severity;
B. Preprocessing
although sensitivity improved in this approach, it still relied on
raw images without any kind of contrast enhancement. The following steps were performed sequentially, after
which preprocessing was complete to enhance quality and
Later works added preprocessing steps to enhance the visibility
converge the model.
of lesions. Lam et al. [3] proposed CLAHE to enhance the
contrast of microaneurysms and hemorrhages, which improved Resizing and Normalization – All images were resized to
the feature extraction capability of CNNs. Similarly, Das et al. 224×224 pixels and pixel values were scaled to the range
[4] applied contrast normalization and resizing before [0,1].
classification with ResNet50; this resulted in higher accuracy for Contrast Enhancement: Contrast Limited Adaptive
early stages of DR. Histogram Equalization (CLAHE) has been used to
More recent work has focused on interpretability of models enhance vessel visibility and bring out microaneurysms.
through explainable AI techniques. Concretely, Selvaraju et al. Noise Reduction – Gaussian blur was applied to suppress
[5] developed the Grad-CAM method that provides good background noise and uneven illumination.
visualization of which image regions drive model predictions,
Data Augmentation: This included random rotation, flip,
thus allowing clinicians to validate model reliability. Along this
zoom, and brightness variations to increase the variety of
line, Mohamed et al. [6] adapted Grad-CAM to highlight lesion
the dataset and prevent overfitting.
areas in DR classification networks, proving that explainability
reinforces clinical trust in AI systems. Train-Validation Split: The data were split into 80% for
training and 20% for validation to assess model
EfficientNet models by Tan and Le [7] have achieved remarkable generalization.
performance with fewer parameters, thus generalizing well. The
work in [8] and [9], among many researchers, adapted it for C. Model Architecture
medical imaging tasks, including DR detection, thanks to its In this work, EfficientNetB0 serves as the base feature
inherent compound scaling technique for efficiency in dealing extractor because it possesses an optimal balance between
with high-resolution images. While most of these studies accuracy and computational efficiency. The architecture
achieved remarkable accuracy, many presented a lack of balance leverages compound scaling in order to uniformly scale
among the three important aspects: accuracy, interpretability, and network width, depth, and resolution. Pretrained weights
computational efficiency. Therefore, the proposed work from ImageNet were utilized for transfer learning of the
combines the EfficientNetB0 architecture with CLAHE network in order to expedite convergence.
preprocessing and integrates the Grad-CAM visualization to
A GAP layer was added after extracting high-level spatial
improve performance and transparency. This approach increases
the reliability for early detection of DR by providing more features from the last convolution block in the
interpretable visual cues for clinical decision-making. EfficientNetB0 to reduce the dimensions. A fully connected
dense layer with ReLU activation was followed by a
III. PROPOSED METHODOLOGY Dropout (0.5) layer to help prevent overfitting. Finally, a
A. Dataset Description softmax output layer classified the image into five classes
for the severity of DR.
Retinal fundus images collected for this study are from
publicly available datasets on diabetic retinopathy, such as Model Layers Overview:
APTOS 2019 and Kaggle Diabetic Retinopathy Detection. Input Layer: 224×224×3
Each image corresponds to one of five DR severity stages: EfficientNetB0 (Base Model, pre-trained)
Global Average Pooling

Dense (512 units, ReLU)

 Dropout (0.5)
Dense (5 units, Softmax)
The model has been compiled with Adam optimizer with
a learning rate of 1e-4 and with categorical cross-entropy
loss for training, using 25 epochs and a batch size of 32
D. Explainability using Grad-CAM
Grad-CAM was integrated into the workflow in order
to enhance interpretability. Grad-CAM uses gradients of
the target class flowing into the final convolutional layer
to produce a heatmap, highlighting image regions that
influenced the model's decision.
This visualization will let the ophthalmologist validate
whether the model is focusing on clinically relevant
features such as microaneurysms, hemorrhages, and
exudates. It provides confidence in model predictions and
supports explainable AI practices in healthcare. In this
work, Grad-CAM was applied to correctly and incorrectly
classified samples for analysis of model attention and
potential sources of error. The generated heatmaps were
found to clearly localize pathological regions in DR
images, thus justifying the reliability of the learning
process.
E. System Architecture Diagram
Fig 1
The proposed system architecture for early detection
[Link] AND DISSCUSION
of DR integrates deep learning-based image classification
with explainable AI visualization. There are five primary A. Model Performance
modules in the workflow: input acquisition, The proposed model for EfficientNetB0 was trained with a
preprocessing, feature extraction, classification, and batch size of 32 on the preprocessed dataset, stopping after
explainability visualization. 25 epochs. This was able to provide for a training accuracy
The proposed overall framework is shown in Fig. 1, with of 96.2% and a validation accuracy of 93.5%, proving good
sequential data flow from the input of retinal images to generalization on unseen data. The loss curves presented
model output and an interpretability layer. smooth convergence without overfitting due to the presence
of dropout along with data [Link] corresponding
Retinal fundus images are collected from publicly
confusion matrix illustrates detection of early-stage DR,
available datasets like APTOS 2019 and Kaggle Diabetic
which comprises mild and moderate conditions with high
Retinopathy datasets. All images are divided into five
precision compared to baseline CNN models. This increase
classes, ranging from No DR to Proliferative DR. Further
in performance is attributed to the compound scaling of
preprocessing includes resizing the images at 224×224
EfficientNetB0 and its ability to extract fine retinal vessel
pixels, followed by CLAHE enhancement for better
features.
vessel visibility, Gaussian filtering for reducing noise,
and normalizing the pixel intensities. To make the data TABLE 1
more robust, rotation, flipping, and brightness Perfomance Metrics
adjustments were performed. Then, the preprocessed
Metrics Training Validation
images will be fed into the EfficientNetB0 network,
which is a lightweight CNN utilized for feature extraction Accuracy
that captures intricate features within the retinas such as Precision
lesions and vessel patterns of the retina. Extracted Recall
features will be put through a classification network with
fully connected layers in which ReLU and Softmax are F1-Score
used as their respective activations. Grad-CAM generates
heatmaps that demarcate those critical retinal regions that These results confirm that the proposed architecture can
contribute most to the class prediction, ensuring the effectively classify fundus images into five classes of DR
model's transparency and interpretability by clinicians.. severity
B. Visualization and Explainability [1] V. Gulshan et al., “Development and Validation of a Deep
. Grad-CAM has been applied to visualize the regions Learning Algorithm for Detection of Diabetic Retinopathy in
influencing the model decision-making process. In this regard, Retinal Fundus Photographs,” JAMA, vol. 316, no. 22, pp. 2402–
the heatmaps for correctly classified images into mild or 2410, 2016.
moderate DR cases included microaneurysms, hemorrhages, [2] H. Pratt et al., “Convolutional Neural Networks for Diabetic
and exudates that are considered the hallmarks of early DR. Retinopathy,” Procedia Computer Science, vol. 90, pp. 200–205,
2016.
In misclassified samples, heatmaps often showed areas with [3] C. Lam et al., “Automated Detection of Diabetic Retinopathy
poor illumination or overlapping lesions; thus, image quality Using Deep Learning,” Ophthalmology, vol. 125, no. 7, pp.
and inconsistent labeling could be causes of poor 1058–1066, 2018.
performance. These visual explanations confirm that the [4] A. Das et al., “Detection of Diabetic Retinopathy Using
model's predictions are based on clinically relevant features ResNet Architecture,” IEEE Access, vol. 8, pp. 193377–193390,
rather than random artifacts. 2020.
By integrating Grad-CAM, the model becomes explainable [5] R. R. Selvaraju et al., “Grad-CAM: Visual Explanations
and interpretable, thus allowing the ophthalmologist to verify from Deep Networks via Gradient-based Localization,” Proc.
predictions and enhancing trustworthiness for clinical IEEE Int. Conf. Comput. Vis. (ICCV), pp. 618–626, 2017.
adoption of the AI. [6] S. Mohamed et al., “Explainable Deep Learning for Diabetic
C. Comparison with Existing Methods Retinopathy Detection Using Grad-CAM,” IEEE Access, vol. 9,
pp. 63456–63465, 2021.
Compared to previous CNN-based models, including
[7] M. Tan and Q. V. Le, “EfficientNet: Rethinking Model
VGG16 and ResNet50, the proposed EfficientNetB0 model
Scaling for Convolutional Neural Networks,” Proc. Int. Conf.
had higher accuracy with fewer parameters. EfficientNetB0's
Mach. Learn. (ICML), pp. 6105–6114, 2019.
compound scaling improved the efficiency of feature
[8] X. Zhang et al., “Diabetic Retinopathy Classification Using
extraction while keeping the architecture lightweight, hence
EfficientNet and Transfer Learning,” IEEE Access, vol. 10, pp.
suitable for use on portable diagnostic devices and
57612–57620, 2022.
teleophthalmology platforms.
[9] P. Kaur et al., “Lightweight CNN Models for Early Stage
TABLE 2 Diabetic Retinopathy Detection,” IEEE Sensors Journal, vol. 23,
Comparison no. 4, pp. 4556–4564, 2023.
Model Parameters Accuracy Explainablity
VGG16 .

ResNet50
Efficient
NetB0

This comparison serves to emphasize the advantage of using

EfficientNetB0 in resource-constrained clinical environments.
D. Conclusion
This paper presents a deep learning approach for the early
detection of Diabetic Retinopathy using EfficientNetB0
combined with Grad-CAM for interpretability. The proposed
model exhibited high accuracy in the classification tasks of
different DR severity levels and showed good performance for
visualizing the disease-affected retinal regions, thereby
reducing the "black box" problem with traditional CNNs. This
work integrates Grad-CAM for improvements in transparency
and clinical trust. EfficientNetB0 assures computational
efficiency. The proposed system can assist the
ophthalmologist as a useful screening tool, especially in rural
or under-resourced areas where specialists may not be readily
available. In the future, this work will be continued by
expanding the dataset with multi-ethnic populations,
incorporating image enhancement techniques, and deploying
the model on a web-based or mobile platform for use in real-
time clinical practice.

V. REFFERENCE