ReXVQA

ReXVQA是一个大规模的视觉问答（VQA）基准数据集，用于胸部X光检查的一般性理解。该数据集包含约696,000个问题，配对160,000个胸部X光研究，跨越训练、验证和测试集。与依赖模板查询的先前工作不同，ReXVQA引入了一个多样化的临床真实任务套件，反映了五个核心放射学推理技能：存在评估、位置分析、否定检测、鉴别诊断和几何推理。数据集来源于四个美国健康系统的胸部X光研究和相应的放射学报告。ReXVQA为评估通用放射学AI系统设立了新的标准，提供公开排行榜、细粒度评估分割、结构化解释和类别级分解。这个基准为下一代AI系统奠定了基础，这些系统能够模仿专家级的临床推理，超越狭窄的病理学分类。

ReXVQA is a large-scale visual question answering (VQA) benchmark dataset for general understanding of chest X-ray examinations. This dataset contains approximately 696,000 questions paired with 160,000 chest X-ray studies, spanning training, validation, and test splits. Unlike prior works that rely on template-based queries, ReXVQA introduces a diverse suite of clinically realistic tasks that embody five core radiological reasoning skills: existence evaluation, localization analysis, negation detection, differential diagnosis, and geometric reasoning. The dataset is sourced from chest X-ray studies and their corresponding radiology reports from four U.S. health systems. ReXVQA sets a new standard for evaluating general-purpose radiological AI systems, offering public leaderboards, fine-grained evaluation splits, structured explanations, and category-level breakdowns. This benchmark lays the foundation for next-generation AI systems that can emulate expert-level clinical reasoning and outperform narrow pathology-specific classification systems.