Surgical foundation model SURGE maps its own limits: safety checks resist pretraining and labels
maier_ak · x · 2026-10-14
Andreas Maier reviews a preprint from Nicolas Padoy's CAMMA group (University of Strasbourg, with TU München, IRCAD and others) measuring where a surgical foundation model stops helping.
Model & data
- SURGE comes in 85.67M and 840.63M parameter sizes, adapted from DINOv3 checkpoints with image- and patch-level self-supervised objectives
- Pretraining corpus SurgSpectrum-30M+ holds 30M+ frames (1 fps) across 132 procedures and 47 anatomical targets; evaluation data excluded, checkpoints released
Findings
- 5 task categories, 15 benchmarks, all via linear probing with supervision scaled from one video upward
- Easy tasks need few labels; reasoning tasks don't benefit much from pretraining: instrument-action-target triplets climb only from 20.30 to 38 mAP over 14 videos
- The clinically critical Critical View of Safety check resists pretraining most; tuning more of the backbone lifts numbers but keeps the shape — "the easy things are solved; the clinically important safety check is not"
Takeaway: label-efficiency gains from foundation models are task-dependent, and surgical safety verification remains unsolved.
Related event: Surgical foundation model SURGE maps the limits of annotation gains(3 posts)→
More from Research
- Neuro for AI & AI for Neuro workshop at AAAI 2027 opens call for papers in Montreal — naturecomputes · 2026-10-14
- BVB Benchmark Tests Agent Video Understanding by Rebuilding 288 Real Videos in Blender — SonglinYang4 · 2026-10-14
- NSF awards $20M to AI Institute in Dynamic Systems for robotic manufacturing — eigensteve · 2026-10-14
- Should ICLR slash its acceptance rate to ~10% as submissions surge? — boknilev · 2026-10-14
- Is 'Attention' the Essence of Intelligence? Quanta examines the idea behind Transformers — TMWNN · 2026-10-14
- A Full QC Pipeline for RL Data: Intake Review, Active Testing, Packaging — joecole · 2026-10-14