DRAN: Densely Reversed Attention based Convolutional Network for Diabetic Retinopathy Detection.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a leading cause of vision loss, diagnosed by analyzing structural factors in fundus images.
• Convolutional Neural Networks (CNNs) are widely used for DR detection, but conventional models may lose spatial correlations.
• Existing multi-stream networks can be computationally expensive.

Purpose of the Study:

• To propose a novel Densely Reversed Attention Network (DRAN) for improved Diabetic Retinopathy detection.
• To leverage multi-level semantic context and channel-wise attention for better spatial representation of DR signs.
• To address the limitations of conventional CNNs in capturing spatial correlations for DR grading.

Main Methods:

• Development of a Densely Reversed Attention Network (DRAN) model.
• Utilizing a pretrained network with learnable integration of channel-wise attention at multi-level features.
• Training and evaluation on the Kaggle DR detection dataset.

Main Results:

• The proposed DRAN model achieved a quadratic weighted kappa of 85.6% on the Kaggle DR detection dataset.
• The approach effectively integrated spatial representations of DR-related factors.
• Performance is competitive with current state-of-the-art methods.

Conclusions:

• The Densely Reversed Attention Network (DRAN) offers a promising approach for accurate Diabetic Retinopathy grading.
• Integrating multi-level features with channel-wise attention enhances the model's ability to capture crucial spatial information.
• DRAN provides a cost-effective and accurate solution for automated DR detection.