Using Deep Learning Architectures for Detection and Classification of Diabetic Retinopathy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a severe complication of diabetes, potentially causing irreversible blindness.
• Early DR detection is critical for timely intervention, but manual grading of retinal images is inefficient and error-prone.
• Automated methods are needed to improve the accuracy and efficiency of DR diagnosis.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for the automated detection and classification of diabetic retinopathy (DR).
• To compare the performance of a hybrid VGG16-XGBoost network with the DenseNet 121 architecture for DR classification.
• To address class imbalance in retinal image datasets for improved model generalizability.

Main Methods:

• Two DL architectures were proposed: a hybrid VGG16-XGBoost classifier and the DenseNet 121 network.
• Retinal images from the APTOS 2019 Blindness Detection Kaggle Dataset were preprocessed and utilized for model training and evaluation.
• Class balancing techniques were applied to mitigate the impact of imbalanced data distribution.

Main Results:

• The hybrid VGG16-XGBoost model achieved an accuracy of 79.50% in DR classification.
• The DenseNet 121 model demonstrated superior performance with an accuracy of 97.30%.
• Comparative analysis confirmed the enhanced efficacy of DenseNet 121 over existing methods on the same dataset.

Conclusions:

• Deep learning architectures, particularly DenseNet 121, show significant potential for accurate and efficient early detection and classification of diabetic retinopathy.
• Automated DR diagnosis systems can enhance diagnostic efficiency and accuracy, benefiting patient outcomes and healthcare systems.
• The DenseNet 121 model represents a promising tool for clinical application in ophthalmology.