Diabetic retinopathy classification network with multi-frequency contextual attention module.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss worldwide.
• Early diagnosis of DR is crucial for timely treatment and preventing blindness.
• Current diagnostic methods face challenges with limited labeled data and subtle lesion identification.

Purpose of the Study:

• To develop an advanced deep learning network, MFCA-DRNet, for accurate diabetic retinopathy classification.
• To overcome limitations of insufficient labeled data and improve the recognition of dispersed lesions in retinal images.

Main Methods:

• A self-supervised contrastive learning strategy was employed for pre-training on the EyePACS dataset.
• An adaptive preprocessing technique combining histogram equalization and non-local means denoising was utilized.
• A multi-frequency contextual attention (MFCA) module and an energy-function-guided attention mechanism were integrated into the network.

Main Results:

• MFCA-DRNet demonstrated strong performance in DR classification across DDR, APTOS 2019, and Messidor-2 datasets.
• The model achieved high accuracy (87.12%), precision (81.2%), recall (85.3%), and F1-score (83.16%) on the APTOS 2019 dataset.
• The MFCA module effectively captured long-range dependencies and enhanced lesion recognition.

Conclusions:

• MFCA-DRNet offers a robust solution for diabetic retinopathy diagnosis, particularly in data-scarce scenarios.
• The proposed methods significantly improve lesion detection and classification accuracy.
• The network shows potential for enhanced clinical applicability in diverse imaging conditions.