Deploying efficient net batch normalizations (BNs) for grading diabetic retinopathy severity levels from fundus...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness globally.
• Early detection through screening is crucial but challenging due to asymptomatic early stages.
• The increasing prevalence of diabetes necessitates efficient DR detection systems.

Purpose of the Study:

• To improve the accuracy and F1 score of deep learning models for diabetic retinopathy detection.
• To leverage pre-trained EfficientNet models with batch normalization (BNs) for feature extraction from fundus images.
• To enhance the performance of these models using data augmentation and Gaussian Smooth filters.

Main Methods:

• Utilized pre-trained EfficientNet BNs for automatic feature extraction from fundus images.
• Applied Gaussian Smooth filters and data augmentation techniques to improve model performance.
• Evaluated the models on the EYE-PACS and DeepDRiD datasets.

Main Results:

• Achieved F1 scores above 80% on the EYE-PACS dataset with all EfficientNet BNs.
• Demonstrated improved accuracy and F1 scores after applying Gaussian Smooth filters and data augmentation.
• Attained final F1 scores of 84% for EYE-PACS and 87% for DeepDRiD.

Conclusions:

• The proposed deep learning approach, incorporating EfficientNet BNs with data augmentation and filtering, significantly enhances diabetic retinopathy detection.
• These improved methods surpass previous study results, offering a more effective tool for early DR diagnosis.
• The findings support the potential of advanced AI techniques in large-scale screening programs to combat DR-related vision loss.