Evidence-Grounded Chest X-ray Report Generation with Retrieval, Citation, and Hallucination Control

Chest X-ray report generation has become an important topic in vision-to-language research. However, fully generative models often create fluent clinical reports that may contain unsupported or inaccurate statements, leading to hallucination problems and reducing reliability. This study investigates evidence-grounded report generation using the Open-i (IU X-ray) dataset with two main goals: generating coherent radiology reports from X-ray images and minimizing unsupported clinical entities through evidence retrieval. Four experimental models were evaluated: a baseline image-to-report model (E1), an alignment-enhanced model using the InfoNCE objective (E2), a retrieval-grounded model that incorporates Top-K evidence sentences with citation markers such as [E#] (E3), and a reranking model that selects the most evidence-supported output (E4). Experimental results on the Open-i test dataset show that grounding methods significantly reduce hallucination rates and improve entity overlap performance. The reranking approach achieves the best grounding quality, although stronger grounding slightly lowers text-overlap scores and increases inference time. Overall, retrieval-based grounding with explicit citations and reranking offers an effective approach for improving factual consistency and reducing unsupported information in automated radiology report generation.