bio-nlp-umass/MedThinkVQA

MedThinkVQA是一个专家标注的基准数据集，专注于放射学中的多图像诊断推理。与以往最多包含单图像的医学视觉问答基准不同，该数据集要求模型从每个图像中提取证据、整合跨视图信息，并进行鉴别诊断推理。它是一个多图像、多模态且部分纵向的基准，包含8,067个案例（7,347个训练案例和720个测试案例），平均每个案例有6.62张图像（测试集为8.12张），涉及CT、MRI、X射线、超声、病理学、临床摄影、核医学与分子成像、内窥镜等多种医学影像模态，30.4%的测试案例具有纵向结构。数据集围绕Think-with-Images工作流设计，提供明确的中间监督，包括三个步骤：单图像发现、集成影像总结和鉴别诊断推理，并包含医学教育案例讨论任务。

MedThinkVQA is an expert-annotated benchmark dataset focused on multi-image diagnostic reasoning in radiology. Unlike previous medical visual question answering (VQA) benchmarks that mostly contain only single images, this dataset requires models to extract evidence from each image, integrate cross-view information, and perform differential diagnostic reasoning. It is a multi-image, multimodal and partially longitudinal benchmark, consisting of 8,067 cases (7,347 training cases and 720 test cases), with an average of 6.62 images per case (8.12 images for the test set), covering various medical imaging modalities including CT, MRI, X-ray, ultrasound, pathology, clinical photography, nuclear medicine and molecular imaging, endoscopy, etc. 30.4% of the test cases have a longitudinal structure. Designed around the Think-with-Images workflow, the dataset provides explicit intermediate supervision, including three steps: single-image discovery, integrated image summarization, and differential diagnostic reasoning, and also includes medical education case discussion tasks.