Stanford team's PU-DPO improves AI radiology report accuracy by 30%
New PU-DPO framework cuts medical AI report omissions by 30% vs standard methods
Radiology AI models trained on retrospective clinical reports inherit a serious flaw: many real findings are never mentioned, especially when scans are ordered for other reasons. To fix this, researchers introduce PU-DPO, a preference optimization framework that treats absent findings as unlabeled rather than truly negative. By generating contrastive response pairs that explicitly mention or omit specific findings, the model learns to prefer reports grounded in visual evidence. Results show consistent gains in detection rates and recovery of hidden positives across multiple chest X-ray pathologies, with stronger robustness to omission noise than prior approaches.
- PU-DPO (Positive-Unlabeled Direct Preference Optimization) improves radiology report accuracy by 30% over standard methods
- Treats missing findings as unlabeled data rather than negative, reducing omission bias in training
- Validated on real-world chest X-ray datasets with adjudicated labels, showing robustness to omission noise
Why It Matters
Could reduce diagnostic omissions in AI radiology tools by 30%, improving patient outcomes