ReXTrust introduces a novel framework that precisely identifies false statements in AI-generated radiology reports by analyzing hidden states in vision-language models.

Key Innovations:

• Produces finding-level hallucination risk scores for precise error detection
• Leverages sequences of hidden states from vision-language models
• Evaluated on the MIMIC-CXR dataset with promising results
• Addresses critical patient safety concerns in medical AI applications

Why This Matters: As radiology increasingly adopts AI-generated reports, robust hallucination detection is essential for maintaining diagnostic accuracy and patient safety. ReXTrust offers a targeted approach to identify potentially harmful errors before they impact clinical decisions.

ReXTrust: A Model for Fine-Grained Hallucination Detection in AI-Generated Radiology Reports