Sahibnoor1/gi-endoscopy-megabank-stage2

GI-Endoscopy MegaBank Stage 2

A large GI/endoscopy image bank prepared for vision-language and encoder training.

Contents

This dataset currently combines:

• HyperKvasir original images
• HyperKvasir video frames extracted at 5 FPS
• Kvasir-Capsule labelled images
• Kvasir-Capsule video frames extracted at 5 FPS

Current size

• Total images/frames: 1,125,053
• HyperKvasir: 283,582
• Kvasir-Capsule: 841,471

Storage format

All images are stored in tar shards for efficient retrieval and training.

Shards

• data/loose_image_shards/: 34 tar shards
• data/kvasir_capsule_video_frame_shards/: 80 tar shards

Metadata files

• metadata/metadata_gi_endoscopy_megabank_stage2_all_sharded.csv — full metadata
• metadata/metadata_hyperkvasir_only.csv — HyperKvasir subset
• metadata/metadata_kvasir_capsule_only.csv — Kvasir-Capsule subset
• metadata/metadata_labelled_only.csv — labelled subset only
• metadata/shard_index.csv — shard-level summary

Main metadata columns

• dataset
• image_id
• storage_type
• shard_file
• path_in_shard
• dataset_section
• origin
• label_status
• label
• source_video
• frame_number
• target_fps

Intended use

• self-supervised encoder training
• SigLIP2 domain adaptation
• GI vision-language research
• weakly supervised and labelled training

Notes

• Some images are labelled, others are unlabelled.
• Kvasir-Capsule video frames are stored as tar shards.
• Use the metadata CSV files to map image_id to shard locations and labels.

Citation / Source

@article{Borgli2020, author = { Borgli, Hanna and Thambawita, Vajira and Smedsrud, Pia H and Hicks, Steven and Jha, Debesh and Eskeland, Sigrun L and Randel, Kristin Ranheim and Pogorelov, Konstantin and Lux, Mathias and Nguyen, Duc Tien Dang and Johansen, Dag and Griwodz, Carsten and Stensland, H{\aa}kon K and Garcia-Ceja, Enrique and Schmidt, Peter T and Hammer, Hugo L and Riegler, Michael A and Halvorsen, P{\aa}l and de Lange, Thomas }, doi = {10.1038/s41597-020-00622-y}, issn = {2052-4463}, journal = {Scientific Data}, number = {1}, pages = {283}, title = {{HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy}}, url = {https://doi.org/10.1038/s41597-020-00622-y}, volume = {7}, year = {2020} }

@article{Smedsrud2021, title = {{Kvasir-Capsule, a video capsule endoscopy dataset}}, author = { Smedsrud, Pia H and Thambawita, Vajira and Hicks, Steven A and Gjestang, Henrik and Nedrejord, Oda Olsen and N{\ae}ss, Espen and Borgli, Hanna and Jha, Debesh and Berstad, Tor Jan Derek and Eskeland, Sigrun L and Lux, Mathias and Espeland, H{\aa}vard and Petlund, Andreas and Nguyen, Duc Tien Dang and Garcia-Ceja, Enrique and Johansen, Dag and Schmidt, Peter T and Toth, Ervin and Hammer, Hugo L and de Lange, Thomas and Riegler, Michael A and Halvorsen, P{\aa}l }, doi = {10.1038/s41597-021-00920-z}, journal = {Scientific Data}, number = {1}, pages = {142}, volume = {8}, year = {2021} }