Large-scale modeling of axonal dynamic responses via deep learning

Area of Science:

• Neuroscience
• Computational Biology
• Biomedical Engineering

Background:

• Large-scale axonal dynamic simulation is crucial for understanding white matter injury but computationally expensive.
• Current methods face significant computational cost, limiting large-scale mechanistic investigations.

Purpose of the Study:

• To develop a computationally efficient method for estimating multimodal axonal injury parameters using deep learning.
• To enable rapid, high-resolution simulations of white matter injury.

Main Methods:

• Trained a convolutional neural network (CNN) using tractography-based fiber strains from head impact simulations.
• Employed a stratified and adaptive sampling strategy to create a minimal yet effective training dataset.
• Validated the CNN's accuracy using independent testing samples, assessing R² and normalized root mean-squared error (NRMSE).

Main Results:

• The CNN achieved high accuracy (R² of 0.91-0.98) and low error (NRMSE of 2.7-5.0%) in predicting axonal injury parameters.
• Demonstrated a 31.5-million-fold efficiency gain compared to conventional direct simulations.
• Successfully generated high-resolution multimodal axonal responses for the entire white matter in seconds.

Conclusions:

• Deep learning, specifically CNNs, offers a powerful solution to overcome computational limitations in white matter injury simulation.
• This approach facilitates large-scale, mechanistic investigations of traumatic brain injury.
• The developed CNN model has the potential to significantly advance future research in neurotrauma and white matter biomechanics.