LVMed-R2 advances medical AI by enhancing large vision-language models to generate more accurate, consistent radiology reports through an innovative perception-reflection reasoning framework.

• Two-Stage Reasoning Process: Combines perception-driven reasoning that extracts key visual findings with reflection-driven reasoning that identifies and corrects logical inconsistencies

• Medical Knowledge Injection: Incorporates domain-specific knowledge through specialized fine-tuning approaches for improved diagnostic accuracy

• Superior Performance: Achieves state-of-the-art results on IU-Xray and MIMIC-CXR datasets, demonstrating significant improvements in both clinical accuracy and report quality

• Practical Clinical Value: Reduces radiologists' reporting workload while maintaining diagnostic reliability, potentially accelerating patient care workflows

LVMed-R2: Perception and Reflection-driven Complex Reasoning for Medical Report Generation