RadThinking
收藏资源简介:
RadThinking是由约翰斯·霍普金斯大学等机构构建的一个用于纵向临床推理的放射学视觉问答数据集。该数据集包含20,362个CT扫描,覆盖9,131名患者,涉及43个癌症组和2,077个经随访验证的健康对照,数据来源于10个欧洲机构2012年至2025年的采集。其创建过程遵循严格的标注协议,由放射科医生进行多阶段肿瘤掩码标注,并与去标识化的放射学报告、临床变量及病理学标签配对。该数据集旨在通过分层问题设计(基础感知、单步推理和组合推理)支持人工智能系统在癌症筛查中进行多步骤临床推理,而非仅仅进行肿瘤检测,应用于医学视觉语言模型训练和强化学习验证。
RadThinking is a radiological visual question answering dataset for longitudinal clinical reasoning, constructed by institutions including Johns Hopkins University. This dataset comprises 20,362 CT scans covering 9,131 patients, involving 43 cancer cohorts and 2,077 healthy controls validated through follow-up. The data was collected from 10 European institutions between 2012 and 2025. Its development follows a rigorous annotation protocol, with multi-stage tumor mask annotations performed by radiologists, paired with de-identified radiological reports, clinical variables and pathological labels. This dataset aims to support AI systems in conducting multi-step clinical reasoning for cancer screening via hierarchical question design (basic perception, single-step reasoning and compositional reasoning), rather than merely tumor detection, and is applied to medical vision-language model training and reinforcement learning validation.
根据您提供的README文件内容,该数据集的详细信息非常有限。以下是基于现有信息的总结:
数据集概述
- 数据集名称:未明确给出(根据网站地址推断为
wenxuanchelsea/RadThinking) - 许可证:Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
- 来源:https://huggingface.co/datasets/wenxuanchelsea/RadThinking
由于README文件仅包含许可证信息,没有提供数据集的描述、用途、组成、规模、使用示例等关键内容,因此无法进一步总结数据集的具体细节。如需更完整的信息,建议直接访问数据集页面或查看其他相关文档。

- 1RadThinking: A Dataset for Longitudinal Clinical Reasoning in Radiology约翰斯·霍普金斯大学·计算机科学系; 巴塞尔大学医院·放射学与核医学诊所; 约翰斯·霍普金斯大学·医学院肿瘤学系 · 2026年



