Lesion-attention pyramid network for diabetic retinopathy grading

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a common complication of diabetes, leading to vision impairment and blindness.
• Automated DR grading systems are crucial for early diagnosis and clinical management.
• Current deep learning methods, while accurate, often lack detailed clinical interpretability.

Purpose of the Study:

• To develop an advanced deep learning model for automated diabetic retinopathy grading.
• To improve the clinical significance of automated DR grading by incorporating lesion-specific information.
• To enhance diagnostic accuracy and provide visual evidence for clinical decision-making.

Main Methods:

• A lesion-attention pyramid network (LAPN) was designed, integrating multi-scale features from different resolution subnetworks.
• A weakly-supervised localization method was employed to generate a lesion activation map from low-resolution features.
• The lesion activation map guided the high-resolution network to focus on relevant lesion areas, with a lesion attention module (LAM) for feature fusion.

Main Results:

• The proposed LAPN demonstrated superior performance compared to existing methods in automated DR grading.
• The model successfully generated a lesion activation map, highlighting consistent lesion regions.
• The lesion activation map served as interpretable evidence, complementing the diagnostic output.

Conclusions:

• The LAPN offers an effective approach for automated diabetic retinopathy grading with enhanced clinical interpretability.
• The method provides lesion-consistent activation maps, aiding ophthalmologists in diagnosis.
• This technology has the potential to significantly improve early detection and management of DR.