TY - JOUR
T1 - FiR-Rad
T2 - Fine-Grained Reinforcement With Structured Reasoning for Chest X-Ray Report Generation
AU - Mei, Xin
AU - Yang, Libin
AU - Gao, Dehong
AU - Cai, Xiaoyan
AU - Han, Junwei
AU - Liu, Tianming
N1 - Publisher Copyright:
© 1982-2012 IEEE.
PY - 2026/6
Y1 - 2026/6
N2 - Automated chest X-ray report generation requires not only clinical accuracy but also transparent and interpretable diagnostic reasoning. In this work, we propose FiR-Rad, a two-stage framework that combines explicit structured reasoning with targeted fine-grained optimization. In the first stage, a supervised chain-of-thought approach guides the model to sequentially analyze and describe a comprehensive range of clinically significant thoracic abnormalities, ensuring clinically meaningful coverage. In the second stage, we introduce a segment-level reinforcement learning strategy based on Group Relative Policy Optimization (GRPO), which assigns precise rewards to each disease-specific reasoning step by evaluating the accuracy of corresponding findings in the synthesized report. This design provides direct feedback for intermediate reasoning and encourages consistency between detailed abnormality analysis and final diagnostic conclusions. Experimental results on the MIMIC-CXR and IU-Xray datasets demonstrate that our framework achieves state-of-the-art performance across clinical and linguistic metrics, with strong zero-shot generalization on IU-Xray. The proposed method significantly enhances interpretability and clinical accuracy, effectively addressing key limitations in automated radiology report generation.
AB - Automated chest X-ray report generation requires not only clinical accuracy but also transparent and interpretable diagnostic reasoning. In this work, we propose FiR-Rad, a two-stage framework that combines explicit structured reasoning with targeted fine-grained optimization. In the first stage, a supervised chain-of-thought approach guides the model to sequentially analyze and describe a comprehensive range of clinically significant thoracic abnormalities, ensuring clinically meaningful coverage. In the second stage, we introduce a segment-level reinforcement learning strategy based on Group Relative Policy Optimization (GRPO), which assigns precise rewards to each disease-specific reasoning step by evaluating the accuracy of corresponding findings in the synthesized report. This design provides direct feedback for intermediate reasoning and encourages consistency between detailed abnormality analysis and final diagnostic conclusions. Experimental results on the MIMIC-CXR and IU-Xray datasets demonstrate that our framework achieves state-of-the-art performance across clinical and linguistic metrics, with strong zero-shot generalization on IU-Xray. The proposed method significantly enhances interpretability and clinical accuracy, effectively addressing key limitations in automated radiology report generation.
KW - Radiology report generation
KW - chain-of-thought
KW - fine-tuning
KW - large multimodal models
KW - reinforcement learning
UR - https://www.scopus.com/pages/publications/105030563651
U2 - 10.1109/TMI.2026.3665561
DO - 10.1109/TMI.2026.3665561
M3 - 文章
C2 - 41701595
AN - SCOPUS:105030563651
SN - 0278-0062
VL - 45
SP - 2922
EP - 2935
JO - IEEE Transactions on Medical Imaging
JF - IEEE Transactions on Medical Imaging
IS - 6
ER -