Replication refined dataset for: A lightweight and extensible cell segmentation and classification model for H&E-stained cancer whole slide images

<p>The refined PanNuke and MoNuSAC Cell Segmentation and Classification Dataset is a unified collection of H&E-stained image patches with cell instance annotations and seven cell-type labels. It is created by combining the PanNuke and MoNuSAC datasets while improving label granularity and consistency across both sources.</p>

<p>The dataset is generated using a cross-relabeling workflow that refines broad or ambiguous classes in each dataset using two ResNet50-based cell classifiers trained on extracted single-cell crops. A classifier trained on MoNuSAC immune cells is used to split the PanNuke inflammatory class into lymphocytes, neutrophils, and macrophages. A classifier trained on PanNuke epithelial subclasses is used to split the MoNuSAC epithelial class into epithelial (benign) and neoplastic (malignant). The relabeled instances are merged with the remaining original classes to form a single dataset with harmonized labels.</p>

<p>The resulting refined dataset includes seven cell types with the following instance counts: neoplastic 105,451; epithelial 29,926; lymphocytes 65,275; neutrophils 3,833; macrophages 3,410; connective 50,585; dead 2,908.</p>