mims-harvard/SPATIA_MIST
收藏资源简介:
MIST(多模态成像和空间转录组学)是一个预训练数据集,用于SPATIA(ICML 2026)。它将细胞级别的基因表达与来自Xenium空间转录组数据的形态学图像裁剪相结合。数据集的构建包括三个阶段:A阶段:从Xenium形态学TIFF图像中裁剪以细胞为中心的图像块,并存储到LMDB数据库中;B阶段:使用本体论元数据注释每个数据集,并将其添加到lamindb集合中;C阶段:将多个每个数据集的LMDB合并为一个文件(可选)。在训练时,数据加载器从lamindb集合读取基因表达,并从LMDB文件读取细胞图像(支持多尺度)。图像在加载时进行预处理:256x256灰度图像转换为RGB堆叠,再通过AutoImageProcessor(ViT-MAE)处理为(3, 224, 224)的浮点张量。
MIST (Multimodal Imaging and Spatial Transcriptomics) is a pre-trained dataset for SPATIA (ICML 2026). It combines cell-level gene expression with morphological image crops derived from Xenium spatial transcriptomics data. The dataset construction consists of three stages: Stage A: Crop cell-centered image patches from Xenium morphological TIFF images and store them in an LMDB database; Stage B: Annotate each dataset with ontological metadata and add it to the lamindb collection; Stage C: Merge multiple LMDB files from individual datasets into a single file (optional). During training, the data loader reads gene expression data from the lamindb collection and cellular images from the LMDB files, which support multi-scale loading. Images are preprocessed at load time: 256×256 grayscale images are converted into RGB stacks, then processed into a (3, 224, 224) floating-point tensor via AutoImageProcessor (ViT-MAE).




