MMRad-IVL-22K
收藏资源简介:
MMRad-IVL-22K是由上海交通大学与浙江大学联合开发的首个大规模胸部X光多模态推理数据集,包含21,994条高质量诊断轨迹。该数据集源自MIMIC-CXR数据库,通过35个解剖区域的系统性标注,构建了视觉证据与文本推理交替的放射科医生工作流。数据生成过程采用三阶段验证框架,结合DeepSeek-v3和Qwen2.5-72B模型的自反思与交叉审计,并由医学专家进行临床保真度评估,最终形成平均2.3步推理链条的解剖学引导多模态数据。该数据集旨在推动医学AI实现放射科医生式的视觉-语言交织推理,解决传统文本链式推理导致的临床幻觉问题,在GPT-5等模型实验中使RadGraph指标提升6%。
MMRad-IVL-22K is the first large-scale chest X-ray multimodal reasoning dataset jointly developed by Shanghai Jiao Tong University and Zhejiang University, containing 21,994 high-quality diagnostic trajectories. Derived from the MIMIC-CXR database, this dataset conducts systematic annotation across 35 anatomical regions to establish a radiologist workflow that alternates between visual evidence and textual reasoning. The data generation process adopts a three-stage validation framework, integrating self-reflection and cross-auditing from DeepSeek-v3 and Qwen2.5-72B models, and is subjected to clinical fidelity evaluation by medical experts, ultimately yielding anatomically guided multimodal data with an average of 2.3-step reasoning chains. This dataset aims to advance medical AI towards radiologist-style visual-language interleaved reasoning, address the clinical hallucination problem caused by traditional textual chain reasoning, and improve the RadGraph metric by 6% in experiments on models such as GPT-5.
- 1Thinking Like a Radiologist: A Dataset for Anatomy-Guided Interleaved Vision Language Reasoning in Chest X-ray Interpretation上海交通大学·计算机学院人工智能研究所; 浙江大学医学院·第一附属医院放射科 · 2026年



