Dataset Card for HyperKvasir

HyperKvasir is the largest publicly available dataset for gastrointestinal (GI) endoscopy, consisting of over 110,000 images and 374 videos, including labeled, unlabeled, and segmented samples. It supports tasks such as classification, segmentation, object detection, and anomaly detection in GI disease diagnostics.

Dataset Details

Dataset Description

HyperKvasir is a comprehensive multi-class dataset collected from real gastro- and colonoscopy examinations at Bærum Hospital, Norway. It contains:

10,662 labeled images across 23 classes
99,417 unlabeled images
1,000 polyp images with segmentation masks and bounding boxes
374 endoscopy videos (11.62 hours)
Curated by: Simula Research Laboratory and Bærum Hospital, Norway
Funded by [optional]: Not specified
Shared by [optional]: Simula Research Laboratory
Language(s) (NLP): English (annotations and metadata)
License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Dataset Sources [optional]

Repository: https://github.com/simula/hyper-kvasir
Paper [optional]: https://doi.org/10.1038/s41597-020-00622-y
Demo [optional]: https://datasets.simula.no/hyper-kvasir/

Uses

Direct Use

Development and benchmarking of computer vision models in medical imaging
Training models for GI tract disease detection
Medical education and training support
Polyp detection and segmentation research

Out-of-Scope Use

Direct clinical diagnosis or patient treatment decisions without expert review
Non-medical imaging tasks or commercial uses without adherence to license terms

Dataset Structure

Labeled Images: 10,662 images in 23 diagnostic and anatomical classes
Unlabeled Images: 99,417 unannotated GI endoscopy images
Segmented Images: 1,000 polyp images with corresponding masks (JPEG) and bounding boxes (JSON)
Videos: 374 AVI format videos with frame-level annotations in 171 sequences

Dataset Creation

Curation Rationale

Created to fill the lack of large, diverse, and annotated GI endoscopy datasets for AI research, especially for deep learning-based detection and classification tasks.

Source Data

Data Collection and Processing

Collected from routine clinical examinations using standard endoscopy equipment. Data were de-identified and curated, with videos annotated by expert gastroenterologists.

Who are the source data producers?

Medical professionals from Bærum Hospital, Norway, including gastroenterologists and technical staff.

Annotations [optional]

Annotation process

Annotations were performed manually by trained gastroenterologists following clinical diagnosis protocols. Segmentations include bounding boxes and pixel-level masks for polyps.

Who are the annotators?

Experienced medical doctors and clinical experts in gastroenterology.

Personal and Sensitive Information

The dataset has been anonymized and de-identified to protect patient privacy. No personally identifiable information is included.

Bias, Risks, and Limitations

Collected from a single hospital, so geographical and device variation is limited.
Some categories may be underrepresented, leading to class imbalance.
Videos are annotated for only a subset of findings.

Recommendations

Users should consider class imbalance and lack of multi-center data when generalizing model results. Medical models trained on this dataset should be validated on diverse clinical data.

Citation [optional]

BibTeX:

@article{borgli2020hyperkvasir,
  title={HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy},
  author={Borgli, Hanna and Thambawita, Vajira and Smedsrud, Pia H and et al.},
  journal={Scientific Data},
  volume={7},
  number={1},
  pages={283},
  year={2020},
  publisher={Nature Publishing Group},
  doi={10.1038/s41597-020-00622-y}
}

APA:

Borgli, H., Thambawita, V., Smedsrud, P. H., Hicks, S., Jha, D., Eskeland, S. L., ... & de Lange, T. (2020). HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific Data, 7(1), 283.

Glossary [optional]

Polyp: Abnormal tissue growths on mucous membranes in the GI tract.
Segmentation mask: A pixel-wise annotation that delineates a specific object or region in an image.
Bounding box: A rectangular box around objects of interest in an image.

More Information [optional]

For further details, visit: https://datasets.simula.no/hyper-kvasir/

Dataset Card Authors [optional]

Simula Research Laboratory team

Dataset Card Contact

[email protected]