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

The Human Protein Atlas project provides 15,050 RGB images of human cells, each resized to 128x128 pixels. These images depict subcellular protein patterns and are labeled for one of 21 organelle localization classes. The dataset was curated by Felix Mohr for the Meta-Album benchmark in June 2022 from a larger multi-label Kaggle competition dataset.

Use Cases

• Train multi-class image classifiers based on 21 distinct protein localization patterns.
• Benchmark model performance on a standardized microscopy dataset with 15,050 images.
• Develop feature extraction methods for small (128x128) biomedical RGB images.
• Study protein organelle distribution in human cells based on labeled microscopy data.

Strengths

• Contains 15,050 labeled images, providing a substantial sample size for model training.
• Images are uniformly preprocessed to a consistent 128x128 pixel RGB format.
• Class labels are derived from the authoritative Human Protein Atlas project.

Limitations

• Dataset was filtered to be multi-class, removing all multi-label images and classes with fewer than 40 members, which may limit certain analyses.
• Column-level documentation is absent; field semantics must be inferred after download.
• The original yellow image channel was dropped during conversion to RGB, which may have removed some signal.

Provenance

Source

The Human Protein Atlas project and the associated Kaggle competition.

Collection Method

Original images were filtered to remove multi-label examples and small classes, converted from RGBY to RGB, and resized.

Freshness

Created on 01 June 2022; last update date is unknown.