EdgeCVS: Democratization of Surgical AI with a Distilled Edge-Deployable Critical View of Safety (CVS) Model

MICCAI 2026 · Amine Yamlahi, Jakob Hennighausen, Pascal Hansen, David Leeb, Lena Maier-Hein

Division of Intelligent Medical Systems (IMSY), German Cancer Research Center (DKFZ), Heidelberg

This release contains the inference code and the EdgeCVS-5M model weights. Training code coming soon.

EdgeCVS distills a 305M-parameter EVA02-Large teacher into a 5M-parameter EdgeNeXt-Small student for Critical View of Safety assessment in laparoscopic cholecystectomy. The teacher pseudo-labels 131K frames taken only from the Calot-triangle-dissection phase. Trained on those frames, the student reaches 66.5 mAP on the SAGES-CVS 2024 test set (teacher: 66.9).

Installation

conda create -n edgecvs python=3.12
conda activate edgecvs
pip install -r requirements.txt

Video inference also needs ffmpeg on the system for H.264 output.

Inference

Run from the repository root. The model is downloaded from Hugging Face on first use.

Folder of images (searched recursively): writes a CSV with image_path and the probabilities C1, C2, C3.

python cls/inference_folder.py input_dir=/path/to/images output_csv=predictions.csv

Video: writes inference_videos/<name>_seg.mp4 (anatomy masks + CVS panel) and <name>_plain.mp4 (CVS panel only). start/end are optional (mm:ss, hh:mm:ss or seconds).

python cls/inference_video.py input_video=/path/to/video.mp4 start=01:10 end=01:25

Add device=cpu to run on the CPU.

License

Code: MIT. Model weights: CC BY-NC 4.0 (trained on data released for non-commercial use).

Citation

@InProceedings{YamAmi_EdgeCVS_MICCAI2026,
        author = { Yamlahi, Amine AND Hennighausen, Jakob AND Hansen, Pascal AND Leeb, David AND Maier-Hein, Lena},
        title = { { EdgeCVS: Democratization of surgical AI with a Distilled Edge-Deployable Critical View of Safety (CVS) model } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16892},
        month = {September},
        page = {pending}
}