Multi-scale multi-attention network for diabetic retinopathy grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) grading is crucial for clinical diagnosis but challenged by intra-class variation and small lesions.
• Deep learning models often struggle to retain information from small lesions and handle feature variability in DR fundus images.

Purpose of the Study:

• To develop a novel multi-scale multi-attention network (MMNet) for improved automatic diabetic retinopathy grading.
• To address the limitations of existing methods in capturing small lesions and diverse feature variations in DR detection.

Main Methods:

• Proposed a lesion attention module combining channel and spatial attention to encode diverse lesion features.
• Introduced a multi-scale feature fusion module to enhance learning from small lesion regions.
• Implemented a Cross-layer Consistency Constraint Loss to mitigate semantic differences across multi-scale features.

Main Results:

• MMNet achieved 86.4% accuracy and 88.4% kappa score for multi-class DR grading on the EyePACS dataset.
• On the Messidor-1 dataset, MMNet reported 98.6% AUC, 95.3% accuracy, 92.7% recall, 95.0% precision, and 93.3% F1-score for referral classification.
• Demonstrated significant improvements over state-of-the-art DR grading methods on two benchmark datasets.

Conclusions:

• MMNet effectively enhances the diagnostic efficiency and accuracy of diabetic retinopathy screening.
• The proposed network advances the application of computer-aided medical diagnosis in DR screening.