AI-Based RGB Image Generation From RG Fundus Images Using Pix2Pix: Validation by Quantitative and Observer-Based...

Area of Science:

• Ophthalmic imaging
• Artificial intelligence in medicine
• Image processing

Background:

• Red-green (RG) fundus images are often captured with older systems.
• Converting RG to red-green-blue (RGB) images can enhance diagnostic utility.
• Generative adversarial networks (GANs) show promise in image synthesis.

Purpose of the Study:

• To evaluate the accuracy and perceptual quality of RGB fundus images generated from RG images using a Pix2Pix-based GAN.
• To assess the feasibility of using AI for fundus image conversion.

Main Methods:

• RG images were extracted from RGB images by disabling the blue laser on an Optos system.
• A Pix2Pix-based conditional GAN was trained for RG to RGB image conversion.
• Quantitative metrics (SSIM, PSNR, LPIPS, MAE, RMSE) and qualitative assessment by 39 ophthalmologists were used.

Main Results:

• Generated RGB images showed high structural similarity (SSIM=0.97) and perceptual quality to original RGB images.
• Ophthalmologists achieved a 57.1% correct classification rate between true and AI-generated images.
• Receiver operating characteristic analysis yielded an area under the curve of 0.497, indicating no significant discrimination.

Conclusions:

• Pix2Pix-based GANs can generate perceptually and structurally consistent RGB images from RG images.
• The AI model does not require lesion-specific attention mechanisms for effective image conversion.
• This technique allows reinterpretation of legacy RG fundus images and supports future AI diagnostic applications.