Related Experiment Videos
Privacy-Preserving Surgical Video Analysis with Swarm Learning - Results from a Multinational Appendectomy Cohort
O L Saldanha1,2, K Pfeiffer1, S Bodenstedt3,4
1Else Kroener Fresenius Center for Digital Health, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany.
NEJM AI
|July 29, 2026
Summary
Artificial intelligence (AI) in surgical video analysis is advancing with weakly supervised deep learning and Swarm Learning. This privacy-preserving method enables multicenter collaboration for improved patient-level outcome prediction from surgical videos.
Area of Science:
- Surgical video analysis
- Artificial intelligence
- Machine learning
Background:
- Current AI in surgical video analysis is limited by manual annotations and data privacy concerns, hindering multicenter collaboration.
- Lack of patient-level outcome data restricts AI model development and validation.
Purpose of the Study:
- To develop and evaluate a pipeline integrating weakly supervised deep learning with Swarm Learning for privacy-preserving, collaborative AI model training on surgical videos.
- To enable patient-level outcome prediction directly from surgical video data, overcoming limitations of manual annotation and data sharing.
Main Methods:
- Developed a pipeline combining weakly supervised deep learning with Swarm Learning, a decentralized approach for collaborative model training without data centralization.
- Evaluated the pipeline on 397 laparoscopic appendectomy recordings from six international centers for appendicitis grading, perforation detection, and inflammation grading.
- Optimized model configurations and compared Swarm Learning performance against single-center and centralized learning approaches.
Main Results:
- The pipeline achieved reliable classification performance for perforation detection using optimal configurations (1 frame/sec, SurgTempoNet).
- Swarm Learning outperformed single-center training and demonstrated performance comparable to centralized learning for both laparoscopic and histopathologic disease staging.
- Key barriers to clinical implementation identified include hardware failure and limited integration with electronic patient records.
Conclusions:
- Weakly supervised deep learning effectively predicts patient-level labels from surgical videos.
- Swarm Learning enables privacy-preserving multicenter collaboration, achieving performance on par with centralized learning for AI in surgical video analysis.
- This approach holds significant potential for advancing clinically relevant, collaborative AI development in surgery.