Hpa10M: 10.5 Million Human Protein Atlas Pathology and Tissue Images

Hpa10M contains 10,495,672 immunohistochemistry (IHC) images derived from the Human Protein Atlas (HPA) by nirschl-lab, updated in February 2026. The collection is split into 7,970,595 pathology images and 2,525,077 tissue images, primarily intended for large-scale medical computer vision research.

Use Cases

• Training deep learning models for pathology image classification using the 7.9M pathology samples
• Developing tissue-type identification algorithms across 2.5M tissue images
• Pre-training large-scale vision transformers on specialized immunohistochemistry (IHC) data

Strengths

• Contains 10,495,672 total images
• Includes 7,970,595 pathology-specific images
• Derived from the authoritative Human Protein Atlas (HPA) source
• Formatted as a WebDataset for efficient large-scale streaming

Limitations

• Extreme imbalance between training (10.4M) and validation (2,000) samples
• High storage and compute requirements due to 10.5 million image records

Provenance

Source

Human Protein Atlas (HPA)

Collection Method

Derived from the Human Protein Atlas repository

Freshness

Last updated February 2026.