DSEGAN: Detail and structure enhanced generative adversarial network for fundus image enhancement

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Fundus imaging is crucial for diagnosing eye diseases, but image quality issues hinder clinical use.
• Supervised methods require paired low- and high-quality images, which are hard to obtain.
• Existing unsupervised methods like CycleGAN struggle with preserving fine details like vessel textures and optic disc contours in fundus images.

Purpose of the Study:

• To propose a novel unsupervised method for enhancing low-quality fundus images.
• To address limitations of existing methods in preserving critical fundus image features.
• To improve the clinical diagnosability and downstream analysis of fundus images.

Main Methods:

• Developed a Detail-Structure Enhanced Generative Adversarial Network (DSEGAN) with a core Detail-Structure Enhanced Generator (DSEGen).
• Incorporated an attention Enhanced Partial Convolution (AEPConv) block for noise suppression and high-frequency detail extraction.
• Utilized a dual-Branch Fusion Block (DBFB) for integrating high-frequency (vessel details) and low-frequency (optic disc structures) information.

Main Results:

• DSEGAN demonstrated superior performance over existing methods on multiple datasets (CFP, UWF, EyeQ, FIVES) based on objective metrics (NIQE, FID, KID).
• The method effectively preserves vessel texture details and optic disc contour structures.
• Enhanced image quality was validated through improved performance in a downstream vessel segmentation task.

Conclusions:

• DSEGAN offers an effective unsupervised approach for fundus image enhancement.
• The proposed AEPConv and DBFB modules are key to preserving crucial image details.
• The enhanced fundus images show improved utility for clinical diagnosis and analysis.