3D Multimodal Grounding Reasoning Dataset
收藏DataCite Commons2025-05-14 更新2025-05-17 收录
下载链接：
https://dataverse.harvard.edu/citation?persistentId=doi:10.7910/DVN/JNGYBC
下载链接
链接失效反馈官方服务：
资源简介：
Our dataset is the first 3D grounded medical image reasoning dataset to date, designed to challenge Vision Language Models in perceiving fine-grained visual details and executing diagnostic reasoning processes analogous to expert clinicians. Focusing on the MRI Osteoarthritis Knee Score (MOAKS), it explicitly models the localized clinical diagnostic workflow through structured chains of thought: anatomical localization, abnormality detection, severity estimation, and final score assignment. We curated 560,000 high-quality quintuples pairing: (1) 3D knee MRI volumes from 2,652 patients, (2) region-specific clinical questions addressing specific MOAKS dimensions, (3) 3D bounding boxes of relevant anatomical structures, (4) expert-crafted diagnostic reasoning steps, and (5) structured MOAKS-based assessments.
提供机构：
Harvard Dataverse
创建时间：
2025-05-14
