orena-SurgZoom

SurgZoom

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

LoRA adapter for Qwen3.5-9B that answers questions about foreign objects in surgical videos. Code: github.com/wxyi057/orena-SurgZoom · Data: HeiCo-FOCUS-VQA · LapChole-FOCUS-VQA

Method

Every question gets 20,480 visual tokens: timestamp and counting questions spend them on 256 frames, all others on 128 sharper frames. A question that names a time is answered on a window around it, and a timestamp answer is refined on ±15 s and ±7.5 s re-cuts. The adapter averages three checkpoints trained on different input regimes.

Results

Usage

git clone https://github.com/wxyi057/orena-SurgZoom && cd orena-SurgZoom
pip install -e .
hf auth login                  # after access to HeiCo-FOCUS-VQA and LapChole-FOCUS-VQA is approved
python tools/make_examples.py
python -m surgzoom.infer --input examples/demo --output answers.json

The surgzoom package reproduces the training-time inputs (frame counts, windows, prompt).

Training

SEGMENT training split of HeiCo-FOCUS-VQA and LapChole-FOCUS-VQA (13,746 questions). LoRA r 64 / α 128 on all linear layers, lr 1e-4, 15 epochs, batch 32, ms-swift 4.3.2. Full recipe: surgzoom_recipe.json.

License

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

Citation

@misc{surgzoom2026,
  title  = {SurgZoom: Question-Guided Temporal Focusing for Surgical Video Question Answering},
  author = {Yi, Weixi and Zhang, Hanyuan and He, Runlong},
  year   = {2026},
  note   = {Solution to the SEGMENT track, ORena SAVE FOCUS Challenge, MICCAI 2026},
  url    = {https://github.com/wxyi057/orena-SurgZoom}
}