Colon Bench: 528 Colonoscopy Videos with Lesion Annotations and Descriptions

528 human-verified colonoscopy videos form a multi-task benchmark for medical vision. It includes over 300,000 bounding boxes, 213,000 segmentation masks, and 133,000 words of clinical descriptions across 14 lesion categories. The dataset was created by ajhamdi and last updated on Hugging Face in March 2026.

Use Cases

• Benchmarking video understanding models based on the 528 annotated colonoscopy videos
• Training lesion detection and classification systems based on the 14 lesion categories including polyps and ulcers
• Developing segmentation models for medical imagery based on the 213,000 provided masks
• Generating or evaluating clinical descriptions based on the 133,000 words of accompanying text

Strengths

• 528 videos provide a substantial base for video-based tasks
• Human-verified annotations increase reliability
• Multi-task design with bounding boxes, segmentation masks, and text descriptions supports diverse model evaluation

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment

Provenance

Source

Hugging Face dataset by author ajhamdi

Collection Method

Likely compiled for research benchmarking, as indicated by the paper and website links

Time Range

null

Freshness

Last updated 2026-03-30 15:51:36

Geography

null