A hybrid neural network approach for classifying diabetic retinopathy subtypes

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy is a leading cause of preventable blindness worldwide.
• Early detection and accurate prediction of diabetic retinopathy are critical for effective patient management.
• Current diagnostic methods may benefit from advanced computational approaches for improved accuracy and speed.

Purpose of the Study:

• To develop and validate a hybrid neural network model for accurate and swift prediction of diabetic retinopathy.
• To enhance the prediction capabilities by integrating EfficientNet and Swin Transformer architectures.
• To improve early diagnosis and treatment planning for diabetic retinopathy patients.

Main Methods:

• A hybrid deep learning model combining EfficientNet and Swin Transformer was developed.
• Local and global features were leveraged to capture intricate lesion characteristics.
• Data augmentation techniques, including Gaussian blur, were employed to optimize model performance.
• Extensive experimental evaluations were conducted on the APTOS 2019 dataset.

Main Results:

• The hybrid model achieved superior performance across all evaluated metrics: sensitivity (0.95), specificity (0.98), accuracy (0.97), and AUC (0.97).
• Significant performance improvements were observed compared to existing mainstream deep learning methods.
• Class activation maps provided interpretable visualizations, aiding in diagnostic decision-making.

Conclusions:

• The proposed hybrid neural network model demonstrates high accuracy and effectiveness in predicting diabetic retinopathy.
• This approach offers robust support for the early diagnosis and timely treatment of diabetic retinopathy.
• The model's interpretability enhances clinical utility for physicians, facilitating better patient care.