Datasets:

nirschl-lab
/

hpa10m

license: cc-by-sa-4.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: hpa10m_train/*.tar
      - split: validation
        path: hpa10m_validation/*.tar
dataset_info:
  features:
    - name: __key__
      dtype: string
    - name: __url__
      dtype: string
    - name: jpg
      dtype: image
    - name: json
      struct:
        - name: comments
          sequence: string
        - name: custom_metadata
          struct:
            - name: area_fraction
              dtype: float64
            - name: area_px
              dtype: int64
            - name: bboxes
              sequence:
                sequence: int64
            - name: caption_1
              dtype: string
            - name: caption_2
              dtype: string
            - name: cell_type
              dtype: string
            - name: ensembl_id
              dtype: string
            - name: file_size_kb
              dtype: float64
            - name: gene
              dtype: string
            - name: generic_caption
              dtype: string
            - name: image_md5
              dtype: string
            - name: patient_age
              dtype: float64
            - name: patient_id
              dtype: int64
            - name: patient_sex
              dtype: string
            - name: rle_mask
              dtype: string
            - name: snomed_code
              dtype: string
            - name: snomed_text
              dtype: string
            - name: staining_intensity
              dtype: string
            - name: staining_location
              dtype: string
            - name: staining_quantity
              dtype: string
            - name: tissue
              dtype: string
            - name: uberon_id
              dtype: string
            - name: uniprot_id
              dtype: string
            - name: url
              dtype: string
        - name: instances
          sequence:
            struct:
              - name: attributes
                sequence: string
              - name: classId
                dtype: int64
              - name: className
                dtype: string
              - name: createdAt
                dtype: string
              - name: creationType
                dtype: string
              - name: error
                dtype: string
              - name: exclude
                sequence: string
              - name: groupId
                dtype: int64
              - name: id
                dtype: string
              - name: locked
                dtype: bool
              - name: pointLabels
                struct:
                  - name: _placeholder
                    dtype: string
              - name: points
                sequence: int64
              - name: probability
                dtype: int64
              - name: type
                dtype: string
              - name: updatedAt
                dtype: string
        - name: metadata
          struct:
            - name: annotatorEmail
              dtype: string
            - name: format
              dtype: string
            - name: height
              dtype: int64
            - name: isPredicted
              dtype: bool
            - name: name
              dtype: string
            - name: pinned
              dtype: bool
            - name: projectId
              dtype: string
            - name: qaEmail
              dtype: string
            - name: status
              dtype: string
            - name: width
              dtype: int64
        - name: tags
          sequence: string

HPA10M Dataset

A large-scale immunohistochemistry (IHC) image dataset derived from the Human Protein Atlas (HPA, https://www.proteinatlas.org/), containing approximately 10.5 million pathology and tissue images with detailed annotations.

Dataset Overview

Statistic	Value
Total Images	10,495,672
Training Set	10,493,672 images (10,497 tar files)
Validation Set	2,000 images (1 tar file)
Image Types	Pathology (7,970,595) / Tissue (2,525,077)
Format	JPEG images + JSON metadata

Directory Structure

hpa10m/
├── README.md                              # This file
├── example_images/                        # Sample images for preview
├── hpa10m_train/                          # Training data (WebDataset tar files)
│   ├── hpa10m_train_0000.tar             # Training shards (10,497 files)
│   ├── hpa10m_train_0001.tar
│   ├── ...
├── hpa10m_validation/                     # Validation data
│   └── hpa10m_validation.tar              # All validation samples (2,000 images)
└── hpa10m_tar_summary/                    # Metadata index files
    └── all.feather                        # Complete index of all images

Data Format

Tar Archives (WebDataset Format)

Each tar file contains paired .jpg and .json files organized by:

Image category: pathology/ or tissue/
Gene prefix: Two-letter gene name prefix (e.g., AB/, CD/)

JSON Metadata Structure

Each image has a corresponding JSON file with rich annotations:

{
  "metadata": {
    "height": 3000,
    "width": 3000,
    "name": "image_filename.jpg",
    "format": ".jpg"
  },
  "custom_metadata": {
    "gene": "TEKT3",
    "ensembl_id": "ENSG00000125409",
    "uniprot_id": "Q9BXF9",
    "tissue": "skin cancer",
    "cell_type": "Tumor cells",
    "patient_id": 3354,
    "patient_age": 92,
    "patient_sex": "male",
    "snomed_code": "M-80703;T-01000",
    "snomed_text": "Squamous cell carcinoma, NOS;Skin",
    "staining_intensity": "negative",
    "staining_location": "none",
    "staining_quantity": "none",
    "generic_caption": "Immunohistochemical staining of human skin cancer...",
    "caption_1": "Detailed caption describing the image...",
    "caption_2": "Alternative caption...",
    "url": "http://images.proteinatlas.org/...",
    "bboxes": [[x, y, w, h], ...],
    "rle_mask": "encoded_segmentation_mask",
    "area_px": 3883806,
    "area_fraction": 0.431534
  }
}

Index Files (Feather Format)

The hpa10m_tar_summary/all.feather file contains an index of all images with columns:

Column	Description
`tar_filename`	Source tar archive name
`split`	Dataset split (train/validation)
`name`	Full path within tar archive
`type`	Image type (pathology/tissue)
`img_offset`	Byte offset of image in tar
`img_size`	Image file size in bytes
`json_offset`	Byte offset of JSON in tar
`json_size`	JSON file size in bytes

Key Annotations

Clinical Information

gene: Gene name (e.g., "TEKT3")
ensembl_id: Ensembl gene ID (e.g., "ENSG00000125409")
uniprot_id: UniProt protein ID (e.g., "Q9BXF9")
tissue: Tissue or cancer type (e.g., "skin cancer")
uberon_id: UBERON ontology ID
cell_type: Cell type (e.g., "Tumor cells")
patient_id: Patient identifier
patient_age: Patient age
patient_sex: Patient sex ("male" / "female")
snomed_code: SNOMED-CT code (e.g., "M-80703;T-01000")
snomed_text: SNOMED-CT description (e.g., "Squamous cell carcinoma, NOS;Skin")

Staining Characteristics

staining_intensity: "negative", "weak", "moderate", "strong"
staining_location: "nuclear", "cytoplasmic/membranous", "cytoplasmic/membranous,nuclear", "none"
staining_quantity: "none", "<25%", "25-75%", ">75%"

Segmentation Data

bboxes: Bounding boxes in [[x, y, width, height], ...] format
rle_mask: Segmentation mask
area_px: Segmented area in pixels
area_fraction: Fraction of image covered by segmentation

Natural Language Captions

generic_caption: Standardized description
caption_1: Detailed scientific description
caption_2: Alternative description

Other Metadata

url: Original image URL from Human Protein Atlas
image_md5: MD5 hash of original image
file_size_kb: Image file size in KB

Usage

Loading Index with Pandas

import pandas as pd

# Load complete index
df = pd.read_feather("hpa10m_tar_summary/all.feather")

# Filter by split
train_df = df[df["split"] == "train"]
val_df = df[df["split"] == "validation"]

# Filter by image type
pathology_df = df[df["type"] == "pathology"]
tissue_df = df[df["type"] == "tissue"]

Data Source

This dataset is derived from the Human Protein Atlas (https://www.proteinatlas.org/), a comprehensive resource for protein expression in human tissues and cancers.

License

Please refer to the Human Protein Atlas data usage terms at https://www.proteinatlas.org/about/licence for licensing information.

📧 Contact

For questions or suggestions, please contact: jjnirschl@wisc.edu or zhi.huang@pennmedicine.upenn.edu