dakomura/tcga-ut

这是一个包含来自32种实体癌类型的组织学图像数据集，总共1,608,060个图像块。数据来源于TCGA数据集，经过两位训练有素的病理学家选择代表性的肿瘤区域并进行质量控制后形成。数据集分为内部和外部两个子集，内部子集确保了训练、验证和测试集之间没有患者重叠，外部子集根据医疗机构分离数据以评估跨机构的泛化能力。数据集可用于31种不同癌症类型的组织学图像分类任务。

This dataset consists of 1,608,060 image patches of hematoxylin & eosin stained histological samples from 32 solid cancer types. The data was sourced from the TCGA dataset, processed through selection of representative tumor regions by two trained pathologists, and quality controlled to remove slides with poor staining or other issues. The dataset is split into internal and external subsets for training, validation, and testing without patient overlap, and to evaluate cross-institutional generalization. It is suitable for the classification task of 31 different cancer types from histopathological images.