MLL23
收藏资源简介:
MLL23数据集包含41906个外周血单个细胞图像,由德国亥姆霍兹慕尼黑研究所的研究人员创建,旨在通过CytoSAE稀疏自编码器模型进行形态学概念发现。该数据集包括来自18种不同细胞类型的细胞图像,覆盖了各种成熟阶段和异常,涵盖了髓系和淋巴系谱系。数据集经过专家标注,用于训练CytoSAE模型,以便在血液学图像中发现具有解释性的形态学特征。CytoSAE模型通过将高维表示分解为稀疏、可解释的组件,实现了细胞形态学概念的学习,并在包括外周血涂片和骨髓细胞学在内的不同数据集上验证了其泛化能力。CytoSAE模型的应用领域包括急性髓系白血病(AML)亚型分类,通过生成患者级别的概念激活和疾病级别的概念分布,揭示了AML亚型的形态学特征,为AI驱动的血液学诊断提供了新的解释性水平。
The MLL23 dataset contains 41,906 single peripheral blood cell images, created by researchers at Helmholtz Munich, Germany, for morphological concept discovery via the CytoSAE sparse autoencoder model. This dataset includes cell images from 18 distinct cell types, covering various maturation stages and abnormalities, and spanning myeloid and lymphoid lineages. The dataset was expertly annotated to train the CytoSAE model for discovering interpretable morphological features in hematological images. The CytoSAE model learns cellular morphological concepts by decomposing high-dimensional representations into sparse, interpretable components, and its generalization ability has been validated across diverse datasets including peripheral blood smears and bone marrow cytology specimens. Applications of the CytoSAE model include acute myeloid leukemia (AML) subtyping: by generating patient-level concept activations and disease-level concept distributions, it reveals the morphological features of AML subtypes, providing a new level of interpretability for AI-driven hematological diagnostics.
CytoSAE: 可解释的血液学细胞嵌入数据集概述
数据集内容
-
训练数据集
- MLL23: 包含41,906张外周血单细胞图像,涵盖18种细胞类型。
- 来源: https://github.com/marrlab/MLL23
- MLL23: 包含41,906张外周血单细胞图像,涵盖18种细胞类型。
-
评估数据集
- Acevedo: 17,092张外周血单细胞图像,标记为11个类别。
- 来源: https://data.mendeley.com/datasets/snkd93bnjr/1
- Matek19: 18,365张专家标注的外周血单细胞图像,分为15个类别。
- 来源: https://doi.org/10.7937/tcia.2019.36f5o9ld
- BMC: 171,373张专家注释的骨髓涂片细胞图像。
- 来源: https://doi.org/10.7937/TCIA.AXH3-T579
- AML_Hehr: 来自189名受试者的患者级单细胞图像,包括四种遗传性AML亚型和对照组。
- 来源: https://doi.org/10.7937/6ppe-4020
- Acevedo: 17,092张外周血单细胞图像,标记为11个类别。
模型权重
- 使用DinoBloom-B嵌入训练CytoSAE模型。
- 模型权重下载地址: https://nefeli.helmholtz-munich.de/records/fdn7v-4vt65/files/final_sparse_autoencoder_dinov2_vitb14_-2_resid_49152.pt?download=1
- 存储路径:
out/checkpoints/8jsxk3co/final_sparse_autoencoder_dinov2_vitb14_-2_resid_49152.pt
相关文件
- 数据集下载与配置指南: DATASET.md
- 训练任务说明: TASKS.md
- 分析演示:
- demo.ipynb
- analysis.ipynb
- patient_analysis.ipynb
参考文献
- Shetab Boushehri等, A large expert-annotated single-cell peripheral blood dataset for hematological disease diagnostics. medRxiv (2025)
- Acevedo等, A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in Brief 30, 105474 (2020)
- Matek等, A single-cell morphological dataset of leukocytes from AML patients and non-malignant controls. (2019)
- Matek等, An expert-annotated dataset of bone marrow cytology in hematologic malignancies. (2021)
- Hehr等, Explainable AI identifies diagnostic cells of genetic AML subtypes. PLOS Digital Health 2(3), e0000187 (2023)
- Koch等, DinoBloom: A foundation model for generalizable cell embeddings in hematology. In: MICCAI (2024)




