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

Romain Hardy, Sung Eun Kim, Du Hyun Ro, Pranav Rajpurkar
Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare, PMLR 281:173-182, 2025.

Abstract

The increasing adoption of AI-generated radiology reports necessitates robust methods for detecting hallucinations—false or unfounded statements that could impact patient care. We present ReXTrust, a novel framework for fine-grained hallucination detection in AI-generated radiology reports. Our approach leverages sequences of hidden states from large vision-language models to produce finding-level hallucination risk scores. We evaluate ReXTrust on a subset of the MIMIC-CXR dataset and demonstrate superior performance compared to existing approaches, achieving an AUROC of 0.8751 across all findings and 0.8963 on clinically significant findings. Our results show that white-box approaches leveraging model hidden states can provide reliable hallucination detection for medical AI systems, potentially improving the safety and reliability of automated radiology reporting.

Cite this Paper

BibTeX
@InProceedings{pmlr-v281-hardy25a, title = {ReXTrust: A Model for Fine-Grained Hallucination Detection in AI-Generated Radiology Reports}, author = {Hardy, Romain and Kim, Sung Eun and Ro, Du Hyun and Rajpurkar, Pranav}, booktitle = {Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare}, pages = {173--182}, year = {2025}, editor = {Wu, Junde and Zhu, Jiayuan and Xu, Min and Jin, Yueming}, volume = {281}, series = {Proceedings of Machine Learning Research}, month = {25 Feb}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v281/main/assets/hardy25a/hardy25a.pdf}, url = {https://proceedings.mlr.press/v281/hardy25a.html}, abstract = {The increasing adoption of AI-generated radiology reports necessitates robust methods for detecting hallucinations—false or unfounded statements that could impact patient care. We present ReXTrust, a novel framework for fine-grained hallucination detection in AI-generated radiology reports. Our approach leverages sequences of hidden states from large vision-language models to produce finding-level hallucination risk scores. We evaluate ReXTrust on a subset of the MIMIC-CXR dataset and demonstrate superior performance compared to existing approaches, achieving an AUROC of 0.8751 across all findings and 0.8963 on clinically significant findings. Our results show that white-box approaches leveraging model hidden states can provide reliable hallucination detection for medical AI systems, potentially improving the safety and reliability of automated radiology reporting.} }
Endnote
%0 Conference Paper %T ReXTrust: A Model for Fine-Grained Hallucination Detection in AI-Generated Radiology Reports %A Romain Hardy %A Sung Eun Kim %A Du Hyun Ro %A Pranav Rajpurkar %B Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare %C Proceedings of Machine Learning Research %D 2025 %E Junde Wu %E Jiayuan Zhu %E Min Xu %E Yueming Jin %F pmlr-v281-hardy25a %I PMLR %P 173--182 %U https://proceedings.mlr.press/v281/hardy25a.html %V 281 %X The increasing adoption of AI-generated radiology reports necessitates robust methods for detecting hallucinations—false or unfounded statements that could impact patient care. We present ReXTrust, a novel framework for fine-grained hallucination detection in AI-generated radiology reports. Our approach leverages sequences of hidden states from large vision-language models to produce finding-level hallucination risk scores. We evaluate ReXTrust on a subset of the MIMIC-CXR dataset and demonstrate superior performance compared to existing approaches, achieving an AUROC of 0.8751 across all findings and 0.8963 on clinically significant findings. Our results show that white-box approaches leveraging model hidden states can provide reliable hallucination detection for medical AI systems, potentially improving the safety and reliability of automated radiology reporting.
APA
Hardy, R., Kim, S.E., Ro, D.H. & Rajpurkar, P.. (2025). ReXTrust: A Model for Fine-Grained Hallucination Detection in AI-Generated Radiology Reports. Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare, in Proceedings of Machine Learning Research 281:173-182 Available from https://proceedings.mlr.press/v281/hardy25a.html.

Related Material