Retinal fundus image super-resolution based on generative adversarial network guided with vascular structure prior

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal fundus image analysis is crucial for diagnosing ophthalmic and systemic diseases.
• Image clarity and resolution are vital for accurate clinical diagnosis.
• Generative adversarial networks (GANs) excel at image super-resolution, but Real-ESRGAN struggles with structural distortions in retinal images.

Purpose of the Study:

• To improve the super-resolution of retinal fundus images by addressing structural distortions.
• To enhance the accuracy and visual quality of retinal images for better disease screening.

Main Methods:

• Utilized a pre-trained U-Net model to generate structural segmentation maps of retinal vessels.
• Integrated spatial feature transform layers to incorporate structural priors into the GAN generator.
• Introduced channel and spatial attention modules into the discriminator's skip connections.
• Incorporated L1 loss on segmentation maps to further constrain the super-resolution process.

Main Results:

• The improved algorithm significantly suppressed structural distortions in super-resolved retinal images.
• Enhanced visual performance and preservation of critical vascular structures were observed.
• Simulation results demonstrated superior quality compared to existing methods.

Conclusions:

• The proposed method effectively overcomes the limitations of existing super-resolution techniques for retinal fundus images.
• This approach offers a promising tool for improving diagnostic accuracy in ophthalmology.
• The integration of structural priors enhances the reliability of deep learning-based image enhancement for medical applications.