Rare diseases have a math problem that no amount of wishful thinking can solve. By definition, they're rare - which means the training data needed to teach AI systems to recognize them is equally scarce. It's the machine learning equivalent of trying to become a bird expert after only seeing three sparrows.
Researchers from multiple institutions just tackled this head-on with EyeDiff, a text-to-image AI that generates synthetic retinal images from written descriptions. Tell it "fundus photograph showing central retinal vein occlusion with flame-shaped hemorrhages," and it produces a medically accurate image that ophthalmologists confirmed looks like the real deal.
The Rare Disease Data Drought
Here's the core challenge: deep learning models are data-hungry beasts. They need thousands, sometimes millions, of examples to learn patterns reliably. But rare eye diseases might only appear in a handful of cases across an entire hospital system. Try training a model on 50 examples when its majority class has 50,000, and you've essentially taught it to ignore rare conditions entirely.
Traditional fixes involve oversampling (copy-pasting those 50 examples repeatedly) or undersampling (throwing away majority class data). Neither is great. Oversampling leads to memorization rather than learning. Undersampling wastes perfectly good data. It's like studying for an exam by either re-reading the same paragraph 100 times or tearing pages out of your textbook.
How EyeDiff Actually Works
The model builds on latent diffusion - the same family of techniques behind image generators you've probably played with. But instead of generating cats in Renaissance paintings, EyeDiff was trained specifically on ophthalmic images: fundus photographs, OCT scans, fluorescein angiography, and 11 other imaging modalities covering over 80 eye diseases.
The training data came from eight large-scale datasets. The model learns to map text descriptions to the statistical patterns that define different retinal conditions - the branching pattern of blood vessels, the fuzzy edges of drusen deposits, the characteristic shapes of hemorrhages. When generating new images, it starts with noise and progressively refines it based on the text prompt, much like a sculptor revealing a form from marble, except the marble is random static and the chisel is gradient descent.
What makes this medically useful rather than just technically impressive is lesion preservation. The generated images don't just look vaguely retinal - they accurately depict the specific pathological features described in the text. Expert ophthalmologists evaluated the outputs and confirmed the synthetic images faithfully represented the clinical findings.
The Results: Actual Numbers
The researchers tested EyeDiff across 11 globally-sourced datasets, using the generated images to augment minority classes before training diagnostic models. The improvements weren't subtle. On the JSIEC dataset, the area under the ROC curve jumped from 0.990 to 0.996, and the precision-recall AUPR rose from 0.887 to 0.967. The Retina dataset saw AUROC improve from 0.857 to 0.892.
More importantly, these gains held across different types of foundation models - whether modality-specific, multimodal, or vision-language models. The synthetic data boosted performance regardless of the underlying architecture, suggesting EyeDiff functions as a general-purpose augmentation tool rather than a quirky trick that only works with specific model families.
Why This Matters Beyond Ophthalmology
The data imbalance problem plagues medical AI everywhere. Rare cancers, unusual presentations of common diseases, conditions that predominantly affect underrepresented populations - all suffer from the same shortage of training examples. If text-to-image generation can reliably produce medically accurate synthetic data, the implications extend well beyond eyes.
The approach also sidesteps some privacy concerns that haunt medical AI development. Synthetic images contain no actual patient data, which could simplify data sharing between institutions and accelerate research collaboration.
Of course, the technique has limits. The model can only generate what it learned to generate - if a particular lesion type or imaging modality wasn't well-represented in training, the synthetic outputs won't magically fill that gap. And generated images, no matter how convincing, require validation against clinical reality before anyone trusts diagnostic models trained on them.
The Bigger Picture
Foundation models in ophthalmology, like RETFound, have already shown that pre-training on large unlabeled datasets creates better starting points for downstream tasks. EyeDiff suggests the next frontier: using generative AI not just to analyze medical images, but to produce the training data that makes analysis possible in the first place.
For the millions of people worldwide affected by rare eye diseases - many of whom face diagnostic delays precisely because their conditions are unfamiliar - better AI tools could mean earlier detection and treatment. The irony isn't lost that teaching machines to recognize rare conditions might require making those conditions less rare, at least in the training data.
References
-
Chen, R., Zhang, W., Liu, B., et al. (2026). Boosting foundation models for rare eye disease diagnosis via a multimodal text-to-image generative framework. npj Digital Medicine. DOI: 10.1038/s41746-026-02560-2
-
Zhou, Y., et al. (2023). A foundation model for generalizable disease detection from retinal images. Nature. DOI: 10.1038/s41586-023-06555-x
-
Hasani, N., et al. (2024). Handling imbalanced medical datasets: review of a decade of research. Artificial Intelligence Review. Link
-
Rajpurkar, P., et al. (2025). Generative Artificial Intelligence in Medical Imaging: Foundations, Progress, and Clinical Translation. PMC. Link
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.