Deep operator network models for predicting post-burn contraction.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Machine Learning

Background:

• Burn injuries are a major global health concern, with contractures causing significant long-term functional impairment.
• Predicting post-burn wound evolution is crucial for effective treatment strategies.
• Traditional finite element models are accurate but computationally intensive, limiting their clinical use.

Purpose of the Study:

• To investigate deep operator networks (a type of neural operator) as a surrogate model for predicting post-burn wound contraction.
• To enhance the deep operator network architecture by incorporating initial wound shape information and boundary condition enforcement.

Main Methods:

• Training a deep operator network on three distinct initial wound shapes.
• Evaluating the network's performance on a test set of wound shapes.
• Comparing the speed and accuracy of the deep operator network against traditional finite element simulations.

Main Results:

• The deep operator network achieved a high R² score of 0.99, demonstrating strong predictive accuracy and generalization.
• The model provided reliable predictions for wound evolution up to one year.
• Significant computational speedups were observed: up to 128x on CPU and 235x on GPU compared to numerical models.

Conclusions:

• Deep operator networks show significant promise as efficient surrogates for finite element methods in simulating post-burn wound evolution.
• This approach has potential applications in accelerating medical treatment planning for burn patients.