A Multi-Degradation Fundus Image Restoration Network Guided by Frequency Prompt.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• High-quality fundus images are essential for diagnosing eye conditions.
• Real-world image acquisition often results in multiple, complex degradations.
• Existing deep learning models struggle to address these multi-component degradations effectively.

Purpose of the Study:

• To develop a unified deep learning framework for restoring fundus images with complex, multi-component degradations.
• To improve the robustness and domain generalization of fundus image restoration models.
• To enhance the clinical utility of fundus images through advanced restoration techniques.

Main Methods:

• Proposing the Multi-degradation Fundus Image Restoration Network (MFR-Net), an all-in-one restoration framework.
• Integrating frequency-aware prompt learning to extract and utilize frequency domain features of degradation components.
• Employing unsupervised domain adaptation in a perceptual and image quality-oriented space for domain alignment.

Main Results:

• MFR-Net demonstrates superior performance compared to state-of-the-art methods in restoring degraded retinal images.
• Significant improvements, up to 5.42%, were observed in quantitative indicators for complex degradations in real-world images.
• The proposed frequency-aware prompt learning and domain adaptation enhance restoration quality and model generalization.

Conclusions:

• MFR-Net offers a comprehensive solution for multi-degradation fundus image restoration.
• The integration of frequency domain features and domain adaptation leads to more effective and generalizable restoration.
• This advancement holds promise for improving diagnostic accuracy in ophthalmology.