An open bone marrow megakaryocyte dataset for automated morphologic studies

Precise classification of megakaryocyte subtypes in bone marrow examination is crucial for the diagnosis and research of various hematological disorders, including Myelodysplastic Syndromes (MDS) and other platelet-production related diseases. While deep learning (DL) has demonstrated remarkable success in medical image classification, its application to megakaryocyte classification has been hindered by the scarcity of high-quality, openly licensed datasets. Therefore, we present MK-11, a dataset comprising 7,204 Wright-Giemsa stained single-cell images across 11 clinically relevant megakaryocyte subtypes. All images were annotated by two experienced hematopathologists with consensus review to ensure annotation quality, following standardized diagnostic criteria. Several state-of-the-art neural networks, including convolutional and transformer-based models, were evaluated on this benchmark, establishing strong baseline performance for megakaryocyte classification. To ensure reproducibility, we provide standardized five-fold cross-validation partitions along with all original images, annotations, partitioning schemes, and evaluation scripts under open licenses. In conclusion, this work built the first public megakaryocyte subtype classification dataset for automatic morphological assessment development and evaluation, serving as a benchmark for future research.骨髓检查中巨核细胞亚型的精准分类，对于包括骨髓增生异常综合征（Myelodysplastic Syndromes, MDS）在内的多种血液系统疾病及血小板生成相关疾病的诊断与研究至关重要。尽管深度学习（Deep Learning, DL）已在医学图像分类领域展现出卓越成效，但其在巨核细胞分类任务中的应用却因高质量开放许可数据集的匮乏而受到掣肘。为此，我们构建了MK-11数据集：该数据集涵盖11种临床相关巨核细胞亚型的7204张瑞氏-吉姆萨染色单细胞图像。所有图像均由两名经验丰富的血液病理学家依据标准化诊断标准完成标注，并经共识审核以保障标注质量。本基准数据集针对多款当前前沿神经网络（包括卷积神经网络及基于Transformer的模型）开展了评估，为巨核细胞分类任务建立了优异的基线性能。为保障研究可重复性，我们以开放许可协议发布了标准化五折交叉验证划分方案，以及全部原始图像、标注文件、划分策略与评估脚本。综上，本研究构建了首个面向自动形态学评估开发与验证的公开巨核细胞亚型分类数据集，可为后续相关研究提供基准参照。