Graph Neural Networks for Realistic Bleeding Prediction in Surgical Simulators

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Physiology

Background:

• Internal bleeding detection is critical in emergency medicine, particularly for trauma cases.
• Existing surgical simulators lack realistic hemorrhage modeling, limiting training effectiveness.
• Accurate prediction of bleeding risk can significantly improve patient outcomes.

Purpose of the Study:

• To develop a novel approach for predicting internal bleeding risk using graph neural networks (GNNs).
• To enhance the realism of surgical simulators by integrating a bleeding risk prediction model.
• To improve emergency response and surgical training for trauma scenarios.

Main Methods:

• Medical images (CT, MRI) were segmented and converted into vascular graph representations.
• A physics-based heuristic (Hagen-Poiseuille equation) was used to calculate bleeding probability for vessel nodes.
• A graph attention network was trained to predict hemorrhage risk from graph data, validated on 1708 vessel graphs.

Main Results:

• The GNN model achieved a mean R-squared of 0.86 (up to 0.9188) in predicting bleeding risk.
• Low mean training (0.0069) and validation (0.0074) losses were observed.
• The model demonstrated robust performance, especially on well-connected vascular graphs, despite data sparsity.

Conclusions:

• Graph neural networks offer a powerful tool for predicting internal bleeding risk from medical imaging.
• The developed model can be integrated into virtual reality surgical simulators for realistic trauma training.
• This approach has the potential to enhance surgical preparedness and improve patient care in emergency situations.