Cross-Center Surgical Step Recognition in Standardized Training Tasks: Dataset, Baselines, and Transfer Analysis

Objective: Accelerating surgical training and improving patient safety requires automated assessment that delivers actionable, step-specific feedback. Standardized box-trainer exercises, with their inherently comparable action steps, offer a natural route towards this goal. Reliable detection of surgical steps is a necessary prerequisite, yet it remains largely unexplored for box-trainer exercises. Crucially, practical deployment requires methods that are invariant to center-domains, i.e. platform, instrument, and camera differences. Methods: We (i) formalize surgical steps for three box-trainer exercises and publicly release the first Multi-center Surgical Training dataset MiST-STEP with step annotations; (ii) benchmark cross-center step recognition with a prevalent two-stage pipeline; (iii) study two self-supervised pretraining schemes—MoCo v2 and a lightweight temporal-order task; and (iv) analyze three finetuning strategies while varying the amount of labeled target center data. Results: Using MiST-STEP (i), we find: (ii) single-center models lose up to 48 percentage points (pp) in macro-F1 on unseen centers; (iii) temporal-order pretraining outperforms MoCo v2 and cuts this deficit by two-thirds; and (iv) finetuning with just two labeled videos per target center removes most of the remaining gap. Conclusion: Our study provides the first comprehensive, cross-center benchmark for step recognition in surgical training and shows that the benefit of self-supervision is highly task-dependent. Our dataset is available at https://gitlab.cc-asp.fraunhofer.de/imte-public/liros/mist. Significance: Center-robust step recognition in box-trainer exercises is essential towards the development of automated feedback systems that aim to standardize training quality and ultimately improve patient outcomes.