Related Experiment Video
Updated: Jan 11, 2026
Robotized Testing of Camera Positions to Determine Ideal Configuration for Stereo 3D Visualization of Open-Heart Surgery
Published on: August 12, 2021
Decentralized, privacy-preserving surgical video analysis with Swarm Learning
Oliver Lester Saldanha1,2, Kevin Pfeiffer1, Sebastian 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.
Artificial intelligence in surgical video analysis is advancing with weakly supervised deep learning and Swarm Learning. This privacy-preserving method enables multicenter collaboration, achieving performance comparable to centralized learning for appendicitis staging.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Surgical Video Analysis
Background:
- Current AI in surgical video analysis is limited by manual annotations and data privacy concerns hindering multicenter collaboration.
- Lack of patient-level outcome integration restricts the clinical utility of AI models.
- Data privacy issues impede the sharing of laparoscopic video data.
Purpose of the Study:
- To develop and evaluate a pipeline integrating weakly supervised deep learning with Swarm Learning for privacy-preserving collaborative surgical video analysis.
- To address limitations of manual annotations and data privacy in AI-driven surgical video analysis.
- To assess the performance of decentralized learning for patient-level disease staging tasks in appendicitis.
Main Methods:
- Developed a pipeline combining weakly supervised deep learning with Swarm Learning, a decentralized approach for collaborative model training without data centralization.
- Utilized a dataset of 397 laparoscopic appendectomy recordings from six international centers.
- Evaluated model performance on binary detection of perforated appendicitis, laparoscopic appendicitis grading, and histopathologic inflammation grading.
Main Results:
- Optimal configurations included 1.0 frames per second sampling and the SurgTempoNet architecture for appendicitis grading.
- Swarm Learning consistently outperformed single-center training and matched centralized learning performance across all disease staging tasks.
- User surveys identified hardware failure and limited electronic health record integration as barriers to clinical implementation.
Conclusions:
- Weakly supervised deep learning allows prediction of patient-level outcomes from surgical videos.
- Swarm Learning enables privacy-preserving multicenter collaboration, achieving performance comparable to centralized learning.
- Decentralized learning holds significant potential for advancing collaborative AI in surgical video analysis, particularly when integrated with electronic health records.
