HyperKvasir: Over 110,000 Gastrointestinal Endoscopy Images and Videos

Over 110,000 images and 374 videos comprise HyperKvasir, the largest publicly available dataset for gastrointestinal endoscopy. It includes labeled, unlabeled, and segmented samples collected from real clinical procedures. The dataset was created by SimulaMet-HOST and was last updated on HuggingFace in May 2025.

Use Cases

• Image classification of gastrointestinal diseases based on labeled endoscopic images.
• Semantic segmentation of anatomical structures and lesions based on segmented samples.
• Object detection for identifying polyps or anomalies in video frames.
• Anomaly detection in gastrointestinal tract imagery using labeled and unlabeled data.

Strengths

• Over 110,000 images provide a substantial scale for training complex models.
• 374 videos offer temporal context for diagnostic tasks.
• Includes labeled, unlabeled, and segmented data, supporting multiple learning paradigms.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for images and videos is not specified, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

SimulaMet-HOST

Collection Method

Collected from real gastrointestinal endoscopy procedures.

Freshness

Last updated 2025-05-20 20:26:49; freshness should be verified.