wxyi088/orena-SurgScope

SurgScope

Our solution to the PROCEDURE track of the ORena SAVE FOCUS Challenge, MICCAI 2026.

LoRA adapter for Qwen3.5-9B that answers questions about foreign objects over entire surgical procedures of up to five hours. Code: github.com/wxyi057/orena-SurgScope · Data: HeiCo-FOCUS-VQA · LapChole-FOCUS-VQA

Method

19 rules on the question text pick the time window. Every question gets ~46k visual tokens: timestamp and counting questions spend them on 1536 frames, all others on 1152 sharper frames. A timestamp answer is re-asked on ±10 min and ±2 min windows around it. The adapter averages three fine-tunes trained with different frame-sampling regimes.

Results

Pre-evaluation	In-distribution	Out-of-distribution
0.6592	0.7195	0.5990

Usage

git clone https://github.com/wxyi057/orena-SurgScope && cd orena-SurgScope
pip install -e .
hf auth login                  # after access to HeiCo-FOCUS-VQA is approved
bash scripts/make_examples.sh
surgscope infer --input examples/test --output answer.json

The surgscope package reproduces the training-time inputs (windows, frame counts, prompt) and merges the adapter into the base model at load time.

Training

PROCEDURE training split of HeiCo-FOCUS-VQA and LapChole-FOCUS-VQA (6,873 questions). LoRA r 64 / α 128 on all linear layers, 15 epochs, ms-swift 4.3.2, three frame-sampling regimes (768; 1536 / 1152; 2880 / 1512 frames). Full recipe: surgscope_recipe.json.

License

CC BY-NC-SA 4.0 (base model: Apache 2.0).

Citation

@misc{surgscope2026,
  title  = {SurgScope: Question-Conditioned Windowing for Hour-Long Surgical Video Question Answering},
  author = {Yi, Weixi and Zhang, Hanyuan and He, Runlong},
  year   = {2026},
  note   = {Solution to the PROCEDURE track, ORena SAVE FOCUS Challenge, MICCAI 2026},
  url    = {https://github.com/wxyi057/orena-SurgScope}
}