| Title |
A Recursive Refinement Framework for Chest X-ray Report Generation |
| Authors |
윤성환(Seonghwan Yoon) ; 임홍기(Hongki Lim) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.6.99 |
| Keywords |
Radiology report generation; Reinforcement learning |
| Abstract |
Generating radiology reports from chest X-rays is challenging due to anatomical overlap and subtle visual evidence. We propose a DPO-based recursive refinement framework that uses special-token logits from a base model trained only on paired image-report data to estimate confidence and iteratively revise sentence candidates, without extra annotations or ROI-based preprocessing. Experiments on MIMIC-CXR show that our method improves clinically oriented evaluation over prior methods. |