Realistic fundus photograph generation for improving automated disease classification.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep convolutional neural networks (CNNs) show promise in classifying retinal diseases.
• Generating realistic medical images is crucial for AI model training.
• Denoising diffusion probabilistic models (DDPMs) are advanced generative models.

Purpose of the Study:

• To assess if DDPMs can create realistic retinal images.
• To evaluate if these generated images can enhance CNN performance for retinal disease classification.
• To compare AI performance against human experts in retinal disease detection.

Main Methods:

• DDPMs were trained to generate retinal fundus images for diabetic retinopathy, age-related macular degeneration, and glaucoma.
• Ophthalmologists evaluated the realism of generated images and performed disease classification.
• Generated images were used to augment training data for CNN ensembles.

Main Results:

• Ophthalmologists achieved 61.1% accuracy in distinguishing real from generated retinal images.
• Augmenting CNN training with 238 generated images significantly improved classification F-score (5.3%) and accuracy (5.8%).
• The enhanced CNN model outperformed the baseline trained solely on real images.

Conclusions:

• DDPMs successfully generate highly realistic retinal images, validated by expert ophthalmologists.
• Incorporating synthetic retinal images into training datasets boosts CNN performance for disease classification.
• This approach enhances diagnostic AI without requiring additional real patient data, addressing data scarcity issues.