Filling in the Missing Piece: Advancing Automated Radiology Report Generation with Clinical Insights

Radiology reports synthesize a radiologist's insights into a patient's condition, serving as vital tools for directing patient care. Despite advances in automated report generation, existing approaches often oversimplify this task, lacking consensus on input selection, ground-truth usage, and output structure. Most methods rely solely on imaging data, merging findings and the impression as a combined output, or incorporate indications as input but generate only the findings. Drawing insights from established medical literature, we clarify the roles and interconnections of report sections and provide definitive answers to some critical questions: (1) What should be the inputs and outputs of a report generation model and why? (2) Can modeling inter-section relationships enhance model performance? We validate our perspective through statistical analyses on MIMIC-CXR, demonstrating universal interdependencies among report sections. We further introduce MIMIC-1V3, a curated dataset derived from MIMIC-CXR where each report is segmented into three sections: the indication, findings, and impression. Furthermore, we propose a R adiologist-L ike P rogressive G eneration (RLPG) framework, explicitly modeling relationships among report sections. RLPG mirrors the real-world clinical workflow in two stages: visual understanding to produce factual observations (findings) of the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$X$</tex>-ray images, and diagnostic reasoning to interpret the implications of findings (impression), addressing the clinical question in the indication.