Challenges in multi-centric generalization: phase and step recognition in Roux-en-Y gastric bypass surgery.

Area of Science:

• Artificial Intelligence in Medicine
• Surgical Workflow Analysis
• Computer Vision in Surgery

Background:

• Current AI models for surgical activity recognition often use limited, single-center datasets, raising concerns about their generalizability.
• Assessing AI model performance across different surgical centers is crucial for real-world clinical applications.

Purpose of the Study:

• To introduce the MultiBypass140, a large, multi-center dataset for laparoscopic Roux-en-Y gastric bypass (LRYGB) surgery.
• To evaluate the generalizability of deep learning models for surgical phase and step recognition using multi-center data.

Main Methods:

• Developed the MultiBypass140 dataset with 140 LRYGB videos from two medical centers (Strasbourg and Bern).
• Annotated surgical phases and steps with input from board-certified surgeons.
• Benchmarked deep learning models through 7 experimental studies, including single-center and multi-center training/evaluation scenarios.

Main Results:

• Models trained on single-center data exhibited limited generalization capabilities.
• Training and evaluating models on the multi-center MultiBypass140 dataset significantly improved generalization.
• Multi-center training enhanced model performance beyond that of independent single-center training and validation.

Conclusions:

• Laparoscopic Roux-en-Y gastric bypass procedures show significant variations in technique and workflow between institutions.
• Multi-center datasets are essential for developing robust AI models that can generalize across different surgical settings.
• The MultiBypass140 dataset and associated code are publicly available to facilitate further research.