emergentai/cervical-cancer-via-images

Cervical Cancer VIA Image Dataset

Overview

This dataset contains pre-stained and post-stained cervical images collected as part of a VIA (Visual Inspection with Acetic Acid) cervical cancer screening workflow. Only image pairs where a trained field nurse and an expert gynecologist reached 100% agreement on the diagnosis were included. This ensures exceptionally high-quality, low-subjectivity ground-truth labels.

The dataset is designed to support research and development of:

• Binary cervical cancer screening/classification models
• Pre/Post–image comparative modeling
• Explainability research in medical imaging
• Clinical decision support systems for low-resource settings

---

Dataset Structure

Each study corresponds to a unique study_id and includes:

• Pre-stained image (before acetic acid application)

• Post-stained image (after acetic acid application)

• Binary VIA label:

  • 0 → Negative
  • 1 → Positive

Metadata file: metadata.csv

Columns:

• study_id
• label_raw ("Positive" or "Negative")
• label (0 or 1)
• pre_local_path (image path inside dataset)
• post_local_path (image path inside dataset)

Image folder:

images/
    <study_id>_pre.jpg
    <study_id>_post.jpg

---

Annotation Process (High-Quality Labels)

All labels in this dataset come from a two-step agreement procedure:

1. Initial annotation by a trained field nurse experienced in VIA screening.
2. Independent review by an expert gynecologist with specialized training in cervical cancer detection.
3. Only samples with 100% agreement between both annotators were uploaded.

This approach:

• Eliminates inter-observer variation
• Provides very clean ground truth
• Ensures the dataset is highly reliable for supervised learning

---

Class Distribution

• Negative: 492 images
• Positive: 47 images

Important: Some image paths were missing in the original storage. Only the successfully copied images are included in this uploaded version.

---

How to Load the Dataset

from datasets import load_dataset

ds = load_dataset("Beijuka/cervical-cancer-via-images")

Each entry includes:

{
  'study_id': 'A003',
  'label': 0,
  'pre_local_path': 'images/A003_pre.jpg',
  'post_local_path': 'images/A003_post.jpg',
  ...
}

---

Ethical Considerations

• Data contains sensitive medical images. Use appropriately.
• Images are anonymized; no personally identifying information is included.
• For research use only — not intended for clinical diagnosis without proper regulatory clearance.
• Redistribution must respect licensing terms.

---