Abstract
Bile duct injury during laparoscopic cholecystectomy is one of the most preventable surgical complications. It can be averted by confirming the Critical View of Safety (CVS), yet this is assessed subjectively and rarely monitored continuously. We investigate whether a single, frozen surgical foundation model already possesses the specialized representations required for automated CVS assessment. Benchmarking five foundation models, we select SurgeNetXL (DINOv3 ViT-L), which nearly doubles the performance of the strongest general domain model. We find that explicit depth and segmentation priors provide no meaningful uplift, consistent with the frozen backbone already encoding much of the relevant 3D geometry (RGB depth CKA = 0.76). Ablating this depth concept via null-space projection yields suggestive evidence that the model selectively uses this implicit geometry: erasing it degrades the 3D structural criterion (C3) while slightly improving the purely 2D semantic criterion (C1). Furthermore, parameter-efficient adaptation (mAP 0.48) outperforms full fine-tuning (0.43) by preventing overfitting. A sparse attention sweep reveals CVS temporal evidence is remarkably short-ranged (2-3 frame optimal causal window). Leveraging this, a confidence-aware causal label propagation framework turns sparse 18-frame anchors into dense targets. Our frozen-feature causal temporal bridge predicts each frame online at mAP 0.49±0.06 up from a 0.41 frozen-probe baseline, at 13.5 ms per frame with latency independent of clip length, outperforming fine-tuned variants and serving as a reliable rule-out signal in our evaluation (NPV ≈99%). As more complex methods do not exceed this bridge, we attribute this ceiling to the sparse-label budget rather than model capacity. Ultimately, strong frozen representations paired with efficient causal reasoning offer a principled path to a deployable CVS safety Monitor.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SafeSurg_020.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/SafeSurg_020_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=Ooh2yrCY0M
BibTex
@InProceedings{ChaSoh_What_MICCAISAT2026,
author = { Chaudhuri, Soham AND Janssen, Gerrit AND Hogerwerf, Remco AND Krol, Joseph AND Manzano Rodríguez, Ana},
title = { { What a Frozen Foundation Model Already Knows about the Critical View of Safety: Toward a Real-Time Monitor } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS pending},
month = {pending},
page = {pending}
}