DYNAMIC-CLEAR: Context-Aware Attribute Weighting for Radiology Report Evaluation

This research introduces DYNAMIC-CLEAR, which adapts CLEAR's attribute weights using contextual cues from referral data or study metadata to better align evaluation metrics with real-world clinical priorities. Unlike CLEAR's static scoring, DYNAMIC-CLEAR integrates a context encoder to generate dynamic weights for five clinically vetted attributes, adjusting emphasis based on clinical scenarios such as emergency cases, oncology follow-ups, or screening studies. A reinforcement learning module calibrates these weights using radiologist feedback to ensure specialty-specific alignment. This approach aims to bridge interpretability gaps, address the limitations of one-size-fits-all evaluation, and enable fairer comparisons of AI-generated radiology reports across diverse clinical settings.

References:

1. CLEAR: A Clinically-Grounded Tabular Framework for Radiology Report Evaluation. Yuyang Jiang, Chacha Chen, Shengyuan Wang, Feng Li, Zecong Tang, Benjamin M. Mervak, Lydia Chelala, Christopher M. Straus, Rêve Chahine, S. Armato, Chenhao Tan (2025). arXiv.org.