Generation of multidisease fundus photographs with code-free platform

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Generating diverse retinal disease fundus images is crucial for training diagnostic models.
• Existing methods often require complex coding techniques.
• High-quality synthetic data can augment real-world datasets for improved AI training.

Purpose of the Study:

• To generate fundus photographs of multiple retinal disease categories using an AI-driven approach.
• To bypass the need for manual coding techniques in image generation.
• To evaluate the utility of synthetic fundus images in clinical diagnostics and AI model training.

Main Methods:

• Utilized the Pix2Pix generative adversarial network (GAN) model on Google Colaboratory.
• Generated synthetic fundus images across 10 categories of retinal conditions.
• Compared diagnostic performance of ophthalmologists and AI models on real versus synthetic images.
• Assessed the ability of humans and AI to differentiate between real and synthetic images.

Main Results:

• Successfully synthesized fundus photographs for 10 retinal disease categories.
• Ophthalmologists demonstrated slightly higher diagnostic accuracy with synthetic images compared to real images.
• Training AI models (ResNet-50, VGG-19) with combined real and synthetic data significantly improved diagnostic accuracy.
• AI image detection websites showed limited ability to distinguish synthetic from real fundus images.

Conclusions:

• Pix2Pix on Google Colaboratory efficiently produces diverse, characteristic fundus images.
• Synthetic fundus images generated by AI show promise for enhancing ophthalmological diagnostics and AI model development.
• The method offers a viable alternative to traditional coding techniques for creating medical imaging datasets.