A Comparative Evaluation of Artificial Intelligence-Based Models in Retinal Image Synthesis

Generative models are increasingly being used to synthesize medical images, with recent advancements in diffusion models demonstrating significant potential for generating realistic imagery. However, generating synthesized retinal fundus images using generative AI models, such as DALL-E 2, has not been thoroughly explored. This study presents a comparative analysis of synthetic retinal image generation between diffusion models and DALL-E 2. First, we trained a Denoising Diffusion Probabilistic Model (DDPM) to generate synthetic images using retinal images from the APTOS 2019 dataset. Next, we employed DALL-E 2 to generate synthetic images based on samples from the same dataset. The generated images were then evaluated by human experts, including three ophthalmologists and two AI experts, and the language model ChatGPT-4, with annotations categorizing each image as real, fake, or uncertain. Our results show that diffusion models consistently produce images with a higher degree of realism, as identified by human experts, whereas images generated by DALL-E 2 are more frequently classified as fake. Notably, this study is the first to conduct a comprehensive evaluation from three perspectives: clinical, technological, and ChatGPT-4-based, comparing diffusion models and DALL-E 2. This interdisciplinary evaluation highlights the potential of diffusion models to improve retinal image synthesis in medical imaging and underscores the importance of multidomain feedback in the development of AI-generated medical images. This study offers a foundation for future research aimed at refining these technologies to enhance their clinical applicability and reliability in medical practice.