Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 21·6 min read
Concept-Guided Noisy Negative Suppression for Zero-Shot Classification and Grounding of Chest X-Ray Findings
TL;DR
Standard contrastive learning for chest X-ray analysis treats images from different patients as negative pairs, but this creates noisy negatives—false negatives where different patients actually share similar findings, causing semantic confusion. The authors developed Concept-Guided Noisy Negative Suppression, a method that identifies and down-weights these misleading negative pairs using medical concepts extracted from radiology reports.
✦ Why It Matters
Engineers building medical imaging systems can adopt this suppression method to train more accurate vision-language models without requiring labeled data.
Key Takeaways
- Identify and suppress misleading negative pairs sharing similar medical concepts during contrastive training.
- Apply concept-guided filtering to improve zero-shot classification without additional labeled data.
- Validate approach on chest X-ray datasets to measure grounding accuracy improvements.
- Consider noisy negatives when designing medical vision-language models for clinical deployment.
Full Summary
Medical imaging AI uses contrastive learning (a training method that learns by comparing similar and dissimilar examples) to align chest X-ray images with radiology reports, enabling zero-shot classification (predicting findings without task-specific training) and grounding (locating findings in images). Standard contrastive learning treats images from different patients as negative pairs, but this creates a fundamental problem: different patients often exhibit identical or similar findings, making these negatives noisy (misleading).
Noisy negatives introduce semantic ambiguity, degrading model performance. The paper proposes Concept-Guided Noisy Negative Suppression, which identifies and suppresses misleading negative pairs by recognizing when different patients share similar radiological concepts.
The method operates during training to reweight or filter negatives based on concept similarity. Results demonstrate improved zero-shot classification accuracy and better localization of findings compared to standard approaches, with implications for deploying medical AI systems that generalize across diverse patient populations.