Machine-Learning-Oncology/orena-segment-annotations

ORena FOCUS 2026 — SEGMENT supplementary annotations (public half)

Supplementary VQA annotations produced by MLO-Lab for the ORena FOCUS 2026 SEGMENT track. This is the openly releasable half; the LapChole-FOCUS half is withheld under that dataset's usage agreement until the organisers publish it.

	rows
vqa/heico_derived.jsonl — HeiCo-FOCUS	1,300
vqa/hernia_mesh.jsonl — hernia videos	140
total	1,440

Also included: raw/hernia_mesh_annotations/ (12 frame-level annotation sheets) and frames/ (2,539 JPEGs, 290 MB).

Frames

HeiCo frames are not redistributed. Every frame is named by its absolute index in the source video, so a question's window plus base_fps (25.0) determines the filenames from your own copy of FOCUS.

Hernia frames are included because the videos are not part of FOCUS and cannot be obtained otherwise:

• frames/clips/<video>/frame<NNNNNNN>.jpg — 2,307 frames the questions refer to. base_fps is 1.0, so index n is second n.
• frames/annotated/<video>/<video>__HH-MM-SS.jpg — 232 frames carrying the sheet annotations.

Every shipped frame lies inside a window that a released annotation covers; no frame outside an annotated segment is included.

Licensing

Not uniform — read before reuse.

• HeiCo-derived questions — derivative annotations of HeiCo-FOCUS; the underlying dataset's terms apply.
• Hernia frames and questions — the source videos carry the standard YouTube licence, not an open licence. We redistribute annotated frames only, never the video, under a direct agreement with the channel owner, for non-commercial research use only. Full provenance with titles and links is in SOURCES.md and sources.json.

Use

Code, schemas and the merge tooling: https://github.com/MLO-lab/segment-orena

python annotations/fetch.py                 # this repo
python segment_track/integrate_released_annotations.py \
    --export sft_export_128 --annotations annotations --n-frames 128

Questions inherit their clip from the matching window in your own official export, so no FOCUS frames need redistributing.

With this half you rebuild 21,440 of the 23,454 training rows: the official 20,000 plus the 1,440 questions here. The remaining 2,014 are the withheld LapChole questions.

Acknowledgements

We are grateful to Dr. Todd S. Harris of California Hernia Specialists, who kindly allowed us to use twelve of his publicly available laparoscopic hernia-repair videos in this work. His permission covers non-commercial research use and the release of the annotated frames; the original videos remain his and are not redistributed. The video titles and links are listed in SOURCES.md.