RadGraph-XL

RadGraph-XL: A Large-Scale Expert-Annotated Dataset for Entity and Relation Extraction from Radiology Reports RadGraph-XL is a large-scale expert-annotated dataset of 2,300 radiology reports sourced from MIMIC-CXR and Stanford Health Care, spanning chest CT, abdomen/pelvis CT, brain MRI, and chest X-ray. Reports were curated through targeted sampling to maximize clinical and semantic diversity, using condition coverage, clustering with sentence embeddings, and length stratification. Each report was double-annotated by board-certified radiologists with entities (anatomy, observations, measurements) and relations (modify, located at, suggestive of), resulting in ~226k entities and ~180k relations. A post-processing pipeline added over 3k measurement entities, supporting quantitative analysis. Inter-annotator agreement was ≥50%, with adjudication resolving disagreements. The dataset is released in structured JSON format, split into training, validation, and test sets, and includes pretrained models and tools for entity–relation extraction. RadGraph-XL sets a benchmark for clinical NLP, enabling research on structured information extraction, measurement understanding, and cross-modality generalization.