JustLLMGRPO Improves Chest X-Ray Generation via Prompt Optimization

Generating realistic chest X-rays from text descriptions is crucial for medical imaging research and applications. Previous efforts primarily focused on improving the image generation models themselves, often treating the input text prompts as static after initial domain adaptation. New research introduces JustLLMGRPO, demonstrating that substantial improvements can be achieved by optimizing the way radiographic information is expressed to an *already adapted* image generator. The method employs an unmodified Large Language Model (LLM) to reformulate prompts, effectively filtering out non-renderable content like temporal comparisons or uncertainties, and emphasizing visible radiographic findings. By applying Group Relative Policy Optimization (GRPO) specifically to the LLM's prompt policy, while keeping the image generator frozen, JustLLMGRPO achieves a significant reduction in image generation errors (RadDINO-FID) and maintains high alignment with the original source prompts. This approach highlights the untapped potential in prompt engineering for specialized generative AI tasks, proving that how information is conveyed can be as critical as the generator's capabilities.