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

The article introduces Med-R2, a novel fine-tuning strategy for automated medical report generation (MRG) that enhances large vision-language models (LVLMs) by incorporating a perception-driven reasoning process and a reflection mechanism. This approach addresses limitations of direct supervised fine-tuning (SFT) by improving the perception of pathological features and integrating radiology-specific knowledge, which leads to increased diagnostic accuracy in generated reports. The proposed method is significant for practitioners as it offers a more robust framework for developing LVLMs that can better interpret medical images and produce accurate reports, thereby improving decision support in healthcare.