Perception-oriented generative adversarial network for retinal fundus image super-resolution.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal fundus imaging is vital for diagnosing eye diseases.
• Image quality is often limited by factors like noise and defocus.
• Existing super-resolution methods can introduce artifacts and lose vessel details.

Purpose of the Study:

• To enhance the perceptual quality of retinal fundus images using a novel generative adversarial network.
• To improve the restoration of fine retinal vessel structures and reduce artifacts.
• To advance super-resolution techniques for better ocular diagnostics.

Main Methods:

• A generative adversarial network (GAN) with a generator and discriminator was developed.
• The generator utilizes multi-scale feature extraction, an image segmentation network, and residual-in-residual dense blocks.
• Attention modules and an artifact loss function were incorporated into the GAN architecture.

Main Results:

• The proposed GAN effectively restored finer details and more accurate retinal vessel structures.
• Generated images exhibited superior perceptual quality compared to state-of-the-art methods.
• Artifacts in texture details were reduced, preserving crucial visual information.

Conclusions:

• The novel GAN significantly improves super-resolution retinal fundus image quality.
• This technique offers enhanced diagnostic potential for early detection of ocular diseases.
• The method provides a robust solution for overcoming limitations in retinal image acquisition.