Group-wise Compression and Summarization via LLM-based Ensemble for Chest X-ray Report Generation

Chest X-ray reports are essential clinical documents that provide detailed analyses of both normal and abnormal areas in X-ray images, offering critical insights for diagnosing a wide range of medical conditions. However, handcrafting X-ray reports is a time-intensive process, requiring radiologists to accurately interpret X-ray images and ensure clinical consistency. To address this challenge, we propose a novel two-step LLM ensemble combined with a disease-based retrieval process designed to enhance the accuracy and reliability of automated X-ray report generation. Our proposed framework first retrieves the most relevant reports by measuring similarity within the image-text embedding space. To ensure clinical relevance, we filter the retrieved reports based on patient-specific disease information. Next, the two-step LLM ensemble refines the filtered reports, progressively summarizing them into a comprehensive summary while preserving key diagnostic insights. This compression process improves coherence, reduces redundancy, and ensures alignment with abnormal findings. Finally, the summarized report, which encapsulates the most meaningful and clinically relevant insights, serves as guidance for generating automated X-ray reports. Our proposed framework surpasses previous methods on widely used benchmarks, including MIMIC-CXR and IU X-ray. The results demonstrate its ability to improve clinical relevance and diagnostic reliability, significantly reducing the workload of radiologists while maintaining high-quality report generation.