PathOPEN / README.md
xuzhl's picture
Use the published dataset ID in the README
f6fd608 verified
|
Raw History Blame Contribute Delete
4.11 kB
metadata
pretty_name: PathOPEN
language:
  - en
license: cc-by-4.0
task_categories:
  - visual-question-answering
tags:
  - histopathology
  - digital-pathology
  - medical-vqa
size_categories:
  - n<1K
configs:
  - config_name: originals
    default: true
    data_files:
      - split: full
        path: data/originals.parquet
  - config_name: cases
    data_files:
      - split: full
        path: data/cases.parquet

PathOPEN

PathOPEN is a dataset for visual question answering in histopathology. Two clinical pathologists selected cases and wrote open-ended, five-option multiple-choice and yes/no questions. This Hugging Face edition repackages the original images, annotations and metadata from PathOPEN v1.0.0 on Zenodo. It preserves the original image bytes and the wording of the questions and answers.

Configurations

Configuration Rows Contents
originals 229 One available original image per row, with its image ID, magnification and the questions, answers and metadata for its case. These images belong to 155 cases.
cases 157 One row per case, with metadata, questions and answers, and the list of collected image slots. This includes cases 9 and 79, whose original images are unavailable.

The case records contain 309 open-ended questions, 157 multiple-choice questions and 157 yes/no questions. An open-ended wrong answer is available for 201 of the 309 questions. Questions apply to all available images in a case and repeat across the corresponding originals rows. Count questions from cases to count each case-level annotation once.

The full Zenodo archive also contains 6,870 augmented images and four manifests for selecting balanced subsets. Download that version when reproducing analyses that use those assets.

Load the original images and case records

from datasets import load_dataset

repo_id = "xuzhl/PathOPEN"
originals = load_dataset(repo_id, "originals", split="full", revision="v1.0.0")
cases = load_dataset(repo_id, "cases", split="full", revision="v1.0.0")

example = originals[0]
image = example["image"]
questions = example["open_ended"]

Pass streaming=True to load_dataset() to iterate through the rows without first downloading the whole configuration.

Fields

  • case_id joins the two configurations. Organ, categorization and regional anatomy retain their stored values.
  • In originals, image holds the original image bytes, image_id identifies the image, and image_slot gives its 1-based position within the case. magnification retains the stored value.
  • In cases, image_slots records the slot number, image ID, original filename and magnification. Missing originals have null image IDs and filenames.
  • open_ended contains the 1-based source question slot, question, correct answer and optional wrong answer. Empty question slots are omitted and missing wrong answers are null.
  • mcq.options preserves the five source options in order. mcq.answer_key retains the source key such as Option 3, which refers to the third option. closed_ended contains the yes/no question and its answer.

The full split contains the full configuration. For model evaluation, assign splits by case_id before selecting images or questions. Keep all images and annotations from each case together. PathOPEN draws on cases from one institution and includes teaching cases, so subgroup results depend on the cases represented.

Licence and citation

The dataset is available under Creative Commons Attribution 4.0 International. Cite the Zenodo version DOI for the dataset and report the Hugging Face configuration if using this edition.

Naeem, A. et al. PathOPEN: A pathologist-curated histopathology dataset for visual question answering. Zenodo (2026), v1.0.0. https://doi.org/10.5281/zenodo.24358341.

The PathOPEN pipeline contains the code used to build the dataset.