Objective:

To address data imbalance in ophthalmic imaging for rare eye diseases using a multimodal generative foundation model that synthesizes clinically realistic images.

Key Findings:

• Synthetic data improved diagnostic accuracy across 11 external datasets, particularly in underrepresented classes.
• Notable improvements in early glaucoma detection, with AUROC rising from 0.860 to 0.927, highlighting the model's effectiveness.
• Generative AI can enhance diagnostic systems in data-scarce subspecialties, addressing the long tail of rare conditions.

Interpretation:

Generative AI, through synthetic data, may help overcome challenges posed by data scarcity in medical diagnostics, particularly for rare conditions.

Limitations:

• Synthetic images may show noticeable differences from real data, such as color and lesion location deviations, potentially introducing diagnostic bias.
• Current training data lacks sufficient population diversity, which may affect the model's generalizability across different demographics.

Conclusion:

Further development is needed to enhance algorithm interpretation and expand datasets for improved diversity and real-time data integration, crucial for effective diagnostics.