Optimizing Diagnostic Text Generation for Medical Imaging with Hybrid LLM Adaptation

This study addresses the challenges faced by large language models (LLMs) in specialized domains such as clinical medicine, including non-professional responses, factual inaccuracies due to insufficient domain knowledge, and high costs associated with data scarcity. To tackle these issues, we constructed a limited-scale manually annotated medical imaging-diagnosis QA dataset and selected the Qwen2.5-7B-Instruct model—which underwent pre-training on a small-scale medical corpus—as the foundational architecture. By integrating parameter-efficient instruction tuning via Low-Rank Adaptation (LoRA) and lightweight reinforcement learning optimization using the Group Relative Policy Optimization (GRPO) algorithm, we propose a two-stage training methodology. Experimental results demonstrate significant improvements in response professionalism and diagnostic accuracy for medical domain tasks. Furthermore, through comparative analysis of different reward mechanisms, we found that BLEU-based reward functions exhibited the most pronounced performance enhancements, particularly achieving a 6.1% increase in ROUGE-1 F1 score compared to the instruction-tuned baseline.