Noah-Kalthoff22's picture
Link Surg-396K to its repo, not the arXiv page, to drop the auto paper association
6c51d7f verified
|
Raw History Blame Contribute Delete
2.98 kB
metadata
license: other
license_name: mixed-see-sources-table
pretty_name: OpScribe GI / Minimally-Invasive Surgical VQA
language:
  - en
task_categories:
  - visual-question-answering
size_categories:
  - 10K<n<100K
annotations_creators:
  - derived
source_datasets:
  - extended
tags:
  - surgical
  - endoscopy
  - laparoscopy
  - robotic-surgery
  - medical
  - vision-language

OpScribe GI / Minimally-Invasive Surgical VQA

No images are included. This is 86,991 question–answer pairs that reference frames in nine public surgical corpora. You obtain the frames yourself, from each source, under that source's own terms.

Built as stage 1 of a training curriculum for SurgVU 2026 Category 2.

Contents

file
data/train.jsonl 75,382 pairs
data/val.jsonl 11,609 pairs
splits.json which case buckets are train and which are val
recipe/build_gi_stage1.py rebuilds the corpus from the source datasets

Each record: question, answer, source, source_image (path relative to that source's own root), case.

Frames were resized to 512×512 for training.

Sources

source pairs procedure link
GraSP / PSI-AVA 30,000 robotic prostatectomy github
CholecT50 15,000 cholecystectomy github
LapGyn4 10,000 gynaecology itec
MultiBypass140 8,198 gastric bypass github
Surg-396K 7,507 mixed (EndoVis-VQLA subset) EndoChat
HeiCo 4,998 colorectal paper
SLAM 4,626 laparoscopic actions figshare
HyperKvasir 3,838 GI endoscopy simula
HeiChole 2,824 cholecystectomy challenge

Please cite all nine.

Licence

Mixed. CholecT50 and HeiCo are CC BY-NC-SA; HyperKvasir is CC BY 4.0; HeiChole requires registration. Check each source before use. We do not relicense any of it.

The split differs from the published one, on purpose

The upstream release split by record, so the same frame lands on both sides: 2,143 images (25.3% of val) also appear in train, and 469 of 615 val directories are on both sides.

This split is by image directory — a real video — hashed into 200 buckets, 180 train and 20 val. Verified: 0 images and 0 directories shared between splits.

Results on this split are not comparable to published numbers on the upstream one. The original files are kept under upstream/ for reference.

Limitations

Stills, not video. Question phrasing varies by source. GraSP is a third of the corpus, so robotic prostatectomy is over-represented.