DennisPolimi/QASNNE

EndoVis18-VQA Out-of-Template

A rephrased version of the EndoVis18-VQA surgical visual question answering benchmark. It tests whether VQA models, and the detectors that flag their unreliable answers, stay stable when the same question is worded differently.

Released with the paper When to Trust the Answer: Question-Aligned Semantic Nearest Neighbor Entropy for Safer Surgical VQA (IJCARS, 2026). Code: DennisPierantozzi/QASNNE.

Files

File	Size	Content
Endovis18-VQA Out-of-Template.zip	2.7 GB	Rephrased (out-of-template) questions; used as the out-of-template split in the paper
EndoVis-18-VQA.zip	2.7 GB	Original EndoVis18-VQA data; used as the in-template split in the paper

Download:

pip install huggingface_hub
huggingface-cli download --repo-type dataset DennisPolimi/QASNNE --local-dir ./data

The data loaders in the code repository (dataloader.py, utils/data_utils.py) read per-sequence annotation files at seq_<N>/vqa/Sentence/*.txt. The evaluation code uses validation sequences 1, 5 and 16 (utils/data_utils.py).

How it was built

Starting from EndoVis18-VQA, the 35 question templates were rephrased. Each rephrasing was drafted with a large language model and then manually verified to preserve the original clinical intent, answer type and referent. Only the wording changes: images, ground-truth answers and train/validation splits are identical to the original.

Intended use

• Stress-testing surgical VQA models under question paraphrase.
• Evaluating failure-detection and uncertainty methods under distribution shift in question wording.

This is a research benchmark for preclinical evaluation. It is not intended for clinical decision-making.

Limitations

• Linguistic variation is limited to the rephrased templates and does not cover the full range of language used in operating rooms; the paper lists richer variation as future work.
• In the paper, failures are defined by a lexical proxy (ROUGE-L below 0.5 against the reference answer), not by expert adjudication.
• Coverage is limited to the procedures, instruments and question types of EndoVis18-VQA.
• Rephrasings come from a single language model and tend to be longer and more explicit than the originals (see the examples above), so they represent one style of variation.

Source and terms

The images and answers come from EndoVis18-VQA (Seenivasan et al., 2022), built on MICCAI EndoVis 2018 videos. Please cite the original dataset and follow its terms of use.

Citation

@article{Carlini2026QASNNE,
  title   = {When to Trust the Answer: Question-Aligned Semantic Nearest Neighbor Entropy for Safer Surgical {VQA}},
  author  = {Carlini, Luca and Pierantozzi, Dennis and Drago, Mauro Orazio and Lena, Chiara and Hassan, Cesare and De Momi, Elena and Stoyanov, Danail and Bano, Sophia and Hoque, Mobarak I.},
  journal = {International Journal of Computer Assisted Radiology and Surgery},
  year    = {2026},
  doi     = {10.1007/s11548-026-03750-9}
}

Original dataset: Seenivasan L, Islam M, Krishna AK, Ren H. Surgical-VQA: Visual Question Answering in Surgical Scenes Using Transformer. MICCAI 2022.