MRG-R3: Retrieval-Reflection-Reward via Multimodal LLMs Advances Clinical Automated Medical Report Generation

Automated medical report generation (MRG) holds promise for tackling mounting radiologist workloads and enhancing diagnostic efficiency. Current approaches, however, are limited by deficiencies in contextual reasoning and clinical integration. Here we introduce MRG-R ³ , a novel Retrieval-Reflection-Reward framework, which enhances MRG by emulating three key behaviors through which radiologists improve diagnostic accuracy. R ³ comprises Retrieval, enabling the model to reference similar past cases during reasoning; iterative Reflection, activated when semantic consistency drops below thresholds; and expert-knowledge-guided reinforcement learning as Reward. Comprehensive evaluation across three datasets over 280k total samples demonstrates that MRG-R ³ achieves superior performance across three groups of nine metrics, substantially outperforming existing state-of-the-art approaches. Radiologists of varying experience consistently prefer MRG-R ³ -generated reports, confirming its alignment with clinical needs. MRG-R ³ , following radiologists' core diagnostic behaviors and empowered by multimodal large language model, establishes a new paradigm for MRG, significantly enhances the clinical validity and linguistic quality of medical reports.