Healthcare imaging
Detecting rare lung pathologies with 4× less real PHI
A radiology AI team used Epineone to expand a thin, consented chest X-ray corpus into a balanced training set covering rare pathologies.
+12.4 pts
AUROC on rare-class subset
4×
less real PHI required
0
patient records re-identifiable
The challenge
Helix's clinical partners could only share 14,000 consented chest X-rays, with severe class imbalance. Rare-but-critical pathologies appeared in fewer than 60 studies — not enough to train a reliable classifier without risking patient privacy by aggregating more real data across hospitals.
How they used Epineone
- Seeded Epineone's medical-imaging generator with the 14,000 consented studies and a structured taxonomy of 23 pathology classes.
- Generated 180,000 synthetic chest X-rays with controlled prevalence, pathology co-occurrence and acquisition-device characteristics.
- Validated synthetic outputs against a held-out real test set using radiologist-graded distributional checks and downstream task accuracy.
- Trained the production model on a 90% synthetic / 10% real mix, with the real data reserved for final fine-tuning and evaluation.
"We finally have a way to teach the model what a rare pathology looks like without waiting two years for enough real cases to accumulate."