ProtoFlow: interpretable and robust surgical workflow modeling with learned dynamic scene graph prototypes

Area of Science:

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

Background:

• AI-assisted surgery requires detailed surgical recognition for advancement.
• High annotation costs, data scarcity, and lack of interpretable models hinder AI progress.
• Scene graphs offer structured surgical event abstraction but are underutilized.

Purpose of the Study:

• Introduce ProtoFlow, a novel framework for interpretable and robust surgical workflow modeling.
• Utilize dynamic scene graph prototypes to capture complex surgical procedures.
• Address limitations in current AI-assisted surgery approaches.

Main Methods:

• Employ a graph neural network (GNN) encoder-decoder architecture.
• Integrate self-supervised pretraining for representation learning.
• Implement a prototype-based fine-tuning stage to discover and refine clinically meaningful patterns.

Main Results:

• ProtoFlow outperforms standard GNN baselines on the CAT-SG dataset.
• Demonstrates exceptional robustness in few-shot learning scenarios, performing well with minimal data.
• Learned prototypes identify distinct surgical sub-techniques and provide interpretable insights into workflow deviations.

Conclusions:

• ProtoFlow advances AI systems for surgery through robust representation learning and explainability.
• Facilitates the development of transparent, reliable, and data-efficient AI tools.
• Accelerates clinical adoption for surgical training, decision support, and workflow optimization.