BACK TO MANUAL / RESEARCHMVM LABS · RESEARCH CASE STUDY // PAPER 001
RESEARCH ARCHIVE // Medical AI · Computer Vision · Deep LearningGLOBAL UNDERGRADUATE AWARDS · HIGHLY COMMENDED · 2025

Early Detection of Diabetic Retinopathy Using EfficientNet-Based Convolutional Neural Networks

ABSTRACT

Diabetic retinopathy is a diabetes-related eye disease and a major cause of preventable blindness, especially in regions with limited access to ophthalmologists. This research developed an EfficientNet-B0-based convolutional neural network to detect and classify diabetic retinopathy into five severity levels: No DR, Mild, Moderate, Severe, and Proliferative. The system used a balanced dataset of 10,000 retinal images, with preprocessing techniques including CLAHE, resizing, and data augmentation to improve image quality and model robustness. Transfer learning and attention-based feature extraction were used to improve diagnostic sensitivity. The model achieved 94.97% accuracy, 94.84% sensitivity, and 95.1% specificity. It was integrated into a Flask-based web application with Grad-CAM visualization, providing interpretable model outputs for physicians and supporting scalable, cost-effective screening in resource-limited healthcare settings.

KEY RESEARCH CONTRIBUTION

An interpretable medical AI screening system that combines EfficientNet-based deep learning, retinal image preprocessing, five-class disease severity classification, and Grad-CAM visualization to support early diabetic retinopathy detection in resource-limited settings.

RESEARCH QUESTIONS
How can deep learning improve early detection of diabetic retinopathy from retinal images?
Can EfficientNet-B0 classify diabetic retinopathy severity levels with high accuracy, sensitivity, and specificity?
How can model interpretability tools like Grad-CAM improve physician trust in AI-assisted diagnosis?
How can medical AI systems be deployed practically in resource-limited healthcare environments?
FIG 01 // SYSTEM MODEL
Retinal ImagePreprocessingCLAHE · ResizeEfficientNet-B0ClassificationGrad-CAM MapPhysician Review
SCHOLARLY SPECS
InstitutionStrathmore University
RegionAfrica & Middle East
AuthorManasseh Maina
Date2025
StatusHighly Commended
KEY RESULTS METRICS
Accuracy94.97%
Sensitivity94.84%
Specificity95.1%
Dataset10,000 images
Classes5 DR Levels
KEYWORDS INDEX
EfficientNet-B0Diabetic RetinopathyMedical ImagingCNNsGrad-CAMFlaskCLAHETransfer Learning
RESEARCH ACTIONS
READ SOURCE PAPER