A Self-Supervised Equivariant Refinement Classification Network for Diabetic Retinopathy Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness, necessitating accurate detection.
• Supervised methods for DR detection require extensive pixel-level annotations, posing a significant challenge.
• Developing automated, annotation-efficient methods for DR classification is crucial.

Purpose of the Study:

• To propose a self-supervised Equivariant Refinement Classification Network (ERCN) for diabetic retinopathy classification.
• To reduce the reliance on large, annotated datasets for accurate DR detection.
• To enhance lesion localization and classification performance in DR detection.

Main Methods:

• Utilized unsupervised contrast pre-training for generalized feature representation learning.
• Implemented self-supervised refinement of class activation maps (CAM) using spatial masking and feature similarity.
• Introduced a hybrid equivariant regularization loss to mitigate CAM refinement issues.
• Employed attention-based multi-instance learning (MIL) for improved feature map instance weighting.

Main Results:

• Achieved 87.4% test accuracy on the EyePACS dataset.
• Attained 88.7% test accuracy on the DAVIS dataset.
• Demonstrated superior performance compared to existing self-supervised DR detection methods.

Conclusions:

• The proposed ERCN method effectively detects diabetic retinopathy with high accuracy.
• Self-supervised learning significantly reduces the need for manual annotations in DR detection.
• ERCN offers a promising, annotation-efficient approach for clinical application in diabetic retinopathy screening.