PanNuke: 19 Human Tissue Images for Histology Classification

PanNuke is a collection of 7,753 histology images across 19 different human tissue types, originally designed for nuclei instance segmentation and classification. The Meta-Album version contains 760 images resized to 128x128 pixels and transformed into a tissue classification task, with stain normalization applied to reduce bias. The dataset was created by Jevgenij Gamper et al. and released in 2019, with a Meta-Album adaptation by Romain Mussard in March 2022.

Use Cases

• Training image classifiers to identify 19 different human tissue types based on the provided labels.
• Developing few-shot learning models for medical image recognition using the standardized Meta-Album format.
• Benchmarking computer vision models on stain-normalized histology images to assess robustness.
• Studying feature representation for nuclei and tissue structures in microscopic imagery.

Strengths

• Contains 760 images specifically formatted for a 19-class tissue classification task.
• Images are preprocessed to a consistent 128x128 pixel size and have undergone stain normalization.
• Derived from a well-cited source dataset (PanNuke) with 7,753 images and 19 tissue types.

Limitations

• The dataset is a transformed subset (760 images) of a larger original collection, which may limit representativeness.
• Column-level documentation and sample data are unavailable, requiring inspection after download.
• The license is non-commercial (CC-BY-NC-SA-4.0), restricting commercial use.

Provenance

Source

PanNuke: An Open Pan-Cancer Histology Dataset for Nuclei Instance Segmentation and Classification by Gamper, Jevgenij et al.

Collection Method

Semi-automatically generated segmentation and classification of nuclei from histology slides, later transformed into a classification task.

Freshness

The Meta-Album version was created on 01 March 2022; the original dataset was published in 2019.