What does this research mean for the field?

An automated deep learning framework for diabetic retinopathy classification achieves high accuracy across multiple datasets, demonstrating feasibility for real-time deployment in telemedicine. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The research aims to improve early detection of diabetic retinopathy using deep learning methods.

March 4, 2026Open Access

Investigations of diabetic retinopathy using deep learning techniques

SWS. Praveen Samuel Washburn GMG. Mahendran

Key Points

The research aims to improve early detection of diabetic retinopathy using deep learning methods.
Developed a deep learning framework for four-stage classification of diabetic retinopathy.
Utilized three benchmark datasets: Messidor, APTOS, and EyePACS.
Applied preprocessing techniques including normalization and contrast enhancement.
Integrated handcrafted GLCM texture descriptors with CNN feature extraction.
Validated real-time deployment using NVIDIA Jetson Nano.
Achieved up to 95% accuracy on the APTOS dataset using DenseNet201.
Demonstrated strong performance across multiple datasets with high precision and recall.
MobileNet outperformed other models like InceptionV3 and VGG16 on the EyePACS dataset.
Confirmed feasibility of lightweight DR screening for telemedicine applications.

Abstract

Diabetes is a chronic metabolic disorder that frequently leads to diabetic retinopathy (DR), a major cause of preventable vision loss worldwide. Early DR detection remains challenging due to the subtle appearance of microaneurysms, haemorrhages, and exudates in retinal fundus images. To address this challenge, this study presents an automated deep learning framework for four-stage DR classification using three benchmark datasets: Messidor, APTOS, and EyePACS. Fundus images were preprocessed using normalization, contrast enhancement, and extensive class-balanced augmentation to improve lesion visibility. A hybrid feature extraction approach was developed by integrating handcrafted gray-level co-occurrence matrix (GLCM) texture descriptors with deep convolutional neural network (CNN) representations to enhance discrimination of fine vascular abnormalities. DenseNet201 and ResNet152 were selected for strong hierarchical feature learning, while MobileNet supported efficient edge deployment. The proposed approach demonstrated strong multi-dataset performance. On the Messidor dataset, the CNN model achieved 91% accuracy, 0.92 precision, 0.89 recall, and a 0.91 F1-score; DenseNet121 achieved 90% accuracy, 0.89 precision, 0.90 recall, and a 0.90 F1-score; and ResNet101 achieved 84% accuracy, 0.90 precision, 0.80 recall, and a 0.85 F1-score. On the APTOS dataset, DenseNet201 achieved 95% accuracy, 0.90 precision, 0.95 recall, and a 0.93 F1-score, CNN achieved 93% accuracy, 0.91 precision, 0.92 recall, and a 0.92 F1-score, and ResNet152 achieved 89% accuracy, 0.87 precision, 0.80 recall, and a 0.84 F1-score. On the EyePACS dataset, MobileNet achieved 89% accuracy, 0.80 precision, 0.87 recall, and a 0.83 F1-score, outperforming InceptionV3 (80% accuracy, 0.78 precision, 0.77 recall, and a 0.78 F1-score) and VGG16 (72% accuracy, 0.70 precision, 0.71 recall, and a 0.71 F1-score). Finally, real-time deployment was validated using an NVIDIA Jetson Nano platform, demonstrating the feasibility of lightweight DR screening in telemedicine and mobile healthcare applications. The integration of GLCM–CNN feature fusion with embedded edge deployment represents a novel contribution not explored in previous DR studies, supporting scalable and accessible DR screening solutions.

Ask AI

Helpful

Bookmark

View Full Paper