Diabetic retinopathy classification using a multi-attention residual refinement architecture.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a leading cause of blindness, necessitating accurate and early detection.
• Conventional Convolutional Neural Networks (CNNs) show promise in DR detection but can be further optimized.
• Enhancing diagnostic feature weighting and spatial information preservation is crucial for improving DR classification.

Purpose of the Study:

• To propose a multi-attention residual refinement architecture for enhancing CNN performance in Diabetic Retinopathy detection.
• To investigate the impact of class-specific multi-attention, space-to-depth preprocessing, and Squeeze-and-Excitation blocks on diagnostic accuracy.
• To demonstrate the framework's versatility across various CNN architectures and its interpretability.

Main Methods:

• Developed a novel multi-attention residual refinement architecture integrating class-specific multi-attention, space-to-depth preprocessing, and Squeeze-and-Excitation blocks.
• Applied the proposed framework to enhance established CNN architectures including ResNet, DenseNet, EfficientNet, and MobileNet.
• Utilized the EyePACS dataset for performance evaluation and generated attention-based visualizations for interpretability.

Main Results:

• Achieved consistent 2-5% performance improvements across multiple CNN architectures on the EyePACS dataset.
• Demonstrated universal applicability and maintained computational efficiency with the proposed enhancements.
• Attention visualizations correlated with known clinical pathological patterns, validating the model's diagnostic reasoning.

Conclusions:

• The multi-attention residual refinement architecture significantly enhances CNN performance for Diabetic Retinopathy detection.
• The framework offers a versatile and interpretable solution applicable to various deep learning models in medical imaging.
• This approach holds potential for improving early DR diagnosis and preventing vision loss in diabetic patients.