Deep learning-enhanced diabetic retinopathy image classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of preventable blindness.
• Early detection and treatment of DR are critical to preserve vision.
• Deep learning (DL) offers potential for automated DR screening.

Purpose of the Study:

• To develop and evaluate a DL model for accurate classification of all five stages of diabetic retinopathy.
• To assess the impact of image augmentation techniques on DL model performance for DR detection.

Main Methods:

• A DenseNet-121 deep learning model was implemented for DR stage identification.
• Two datasets, APTOS and DDR, were utilized for model training and validation.
• Image augmentation techniques were applied to enhance dataset quality and balance.

Main Results:

• The proposed DL model achieved high accuracy on both datasets: 98.36% test accuracy on APTOS and 79.67% on DDR.
• Top-2 and Top-3 accuracy scores exceeded 92% and 98% on the DDR dataset.
• Precision, recall, and F1-score metrics confirmed the model's efficacy in DR stage classification.

Conclusions:

• Deep learning models can effectively identify multiple stages of diabetic retinopathy.
• Utilizing higher-quality, augmented images significantly enhances DL model performance.
• This approach holds promise for improving DR screening efficiency and accuracy.