Meta-Album PRT: Subcellular Human Protein Images for Multi-Class Classification

800 RGB images of human cells, each labeled with one of 20 protein organelle localizations. This dataset is a curated subset of the Human Protein Atlas, originally from a Kaggle competition, where multi-label images were filtered and resized to 128x128 pixels. The data was released under CC BY-SA 3.0 by the Meta-Album project in June 2022.

Use Cases

• Train multi-class image classifiers based on 20 distinct protein localization patterns.
• Benchmark model performance on a standardized microscopy dataset with 800 images.
• Fine-tune pre-trained models for biological image recognition tasks.
• Study feature representation for subcellular structures in resized 128x128 RGB images.

Strengths

• Contains 800 images with clear, single-class labels across 20 categories.
• Images are uniformly preprocessed to a standard 128x128 pixel RGB format.
• Derived from the authoritative Human Protein Atlas project, with clear provenance and citation.

Limitations

• Dataset is a filtered subset; original multi-label information and images from rare classes (under 40 members) were removed.
• Image resolution is reduced from the original 512x512 pixels, which may limit analysis of fine details.
• Row count and column-level documentation are unknown, requiring inspection after download.

Provenance

Source

The Human Protein Atlas / Kaggle competition 'Human Protein Atlas Image Classification'.

Collection Method

Original RGBY images were filtered to single-class examples, the yellow channel was dropped, and images were resized.