Instantaneous Brain Strain Estimation for Automotive Head Impacts via Deep Learning

Area of Science:

• Biomechanics
• Computational Neuroscience
• Injury Prevention

Background:

• Accurate brain strain estimation is crucial for head injury modeling in automotive impacts.
• Existing methods struggle with complex, long-duration impact profiles.
• Convolutional Neural Networks (CNNs) show promise for rapid, detailed strain analysis.

Purpose of the Study:

• To extend CNN application for estimating brain strains in automotive head impacts.
• To evaluate different CNN training strategies for optimal performance.
• To assess the CNN's accuracy against established injury metrics.

Main Methods:

• Utilized head impact kinematics from public databases, augmented for increased dataset size.
• Simulated impacts using the Worcester Head Injury Model (WHIM) to generate peak maximum principal strain (MPS) data.
• Employed concatenated rotational velocity and acceleration profiles as CNN input, comparing baseline, transfer learning, and combined training strategies.

Main Results:

• Combined training strategy yielded the best performance, achieving R²=0.932 and RMSE=0.031 for peak MPS on real-world data.
• CNN demonstrated high accuracy for elementwise MPS and Cumulative Strain Damage Measure (CSDM) estimations.
• Validated CNN performance on independent car crash impact data, showing strong correlation coefficients (k=0.98, r=0.90).

Conclusions:

• The developed CNN efficiently and accurately estimates elementwise brain strains in automotive head impacts.
• This CNN approach surpasses conventional kinematic injury metrics in detail and accuracy.
• The CNN technique holds significant potential for improving head protective countermeasure design in the automotive industry.