Fig. 2: Representative example of the RaR process for a radiology question answering item.
From: Multi-step retrieval and reasoning improves radiology question answering with large language models
This figure shows the full RaR workflow for a representative question (RSNA-RadioQA-Q53) involving a patient with systemic symptoms and a low signal intensity left atrial mass associated with the interatrial septum. The pipeline begins with keyword-based summarization to guide retrieval, followed by parallel evidence searches for each diagnostic option using Radiopaedia.org. Retrieved content is synthesized into a structured report, including an introduction, citation-backed analyses of all options (cardiac myxoma, papillary fibroelastoma, rhabdomyosarcoma, and left atrial thrombus), and a neutral conclusion. The approach supports interpretable, evidence-grounded radiology question answering.
