Search papers, labs, and topics across Lattice.
The paper introduces Evidence-aware Self-Correcting Reinforcement Learning (ESC-RL) to improve radiology report generation by addressing limitations in evidence grounding and alignment with clinical preferences. ESC-RL uses a Group-wise Evidence-aware Alignment Reward (GEAR) to provide granular feedback based on true positives, false negatives, and false positives, and a Self-correcting Preference Learning (SPL) strategy to refine reports using an LLM and a disease-aware preference dataset. Experiments on chest X-ray datasets show ESC-RL achieves state-of-the-art performance and improved clinical faithfulness.
LLMs can now autonomously refine their radiology report generation skills by learning from their own mistakes, without human intervention.
Recent reinforcement learning (RL) approaches have advanced radiology report generation (RRG), yet two core limitations persist: (1) report-level rewards offer limited evidence-grounded guidance for clinical faithfulness; and (2) current methods lack an explicit self-improving mechanism to align with clinical preference. We introduce clinically aligned Evidence-aware Self-Correcting Reinforcement Learning (ESC-RL), comprising two key components. First, a Group-wise Evidence-aware Alignment Reward (GEAR) delivers group-wise, evidence-aware feedback. GEAR reinforces consistent grounding for true positives, recovers missed findings for false negatives, and suppresses unsupported content for false positives. Second, a Self-correcting Preference Learning (SPL) strategy automatically constructs a reliable, disease-aware preference dataset from multiple noisy observations and leverages an LLM to synthesize refined reports without human supervision. ESC-RL promotes clinically faithful, disease-aligned reward and supports continual self-improvement during training. Extensive experiments on two public chest X-ray datasets demonstrate consistent gains and state-of-the-art performance.
