K7 - Rapid Prediction of Multimodal Axonal Responses Using a Deep Learning Model

Introduction: : Traumatic axonal injury (TAI) is the hallmark of traumatic brain injury (TBI), which involves axonal damage potentially throughout the white matter. Traditional finite elements (FE) models capture brain tissue deformation at the organ level but cannot resolve subcellular damage within individual axons. Conversely, microscale axon models can simulate the mechanics of axonal injury in detail, but these simulations are computationally prohibitive and cannot feasibly be extended to the millions of axons in the whole white matter. Therefore, while it is valuable to evaluate microscale axonal injury across the entire white matter, it is also extremely challenging due to the computational bottleneck. To resolve this limitation, we were inspired by the recent success of using a deep learning surrogate to dramatically improve the model simulation efficiency compared to the conventional direct model simulation using a whole-brain injury model. Here, we similarly develop a deep learning approach using a convolutional neural network (CNN) to rapidly inference multimodal axonal responses. The CNN is trained to learn the complex relationship between fiber strain profiles and microscopic axonal responses, enabling rapid (on the order of seconds) and accurate predictions of axonal responses throughout the white matter. This model allows future large-scale investigations of axonal responses not feasible with conventional, direct simulation methods.

Materials and

Methods: : We built upon an established voxelization pipeline to generate fiber strain (FS) profiles for a concussed male ice-hockey player, using head impact data recorded by an instrumented mouthguard (iMG) (Zhang et al., 2024). Subject-specific finite element simulations yielded millions of tractography-based FS profiles. A previously developed male axonal injury model (Zhang et al., 2023) was used to simulate four injury responses: peak maximum principal strain (MPS) in microtubules (MT) and axolemma, and failure percentages of tau proteins and neurofilaments (NFs).
Of the 47 recorded head impacts, 46 were used for CNN training and one was reserved for independent testing. To reduce the computational cost required for generating training samples, we applied a stratified, adaptive sampling strategy. First, FS profiles with peak magnitudes >10% were selected. These were further refined using Euclidean distance-based similarity scores to ensure diversity in samples. In the end, we selected 2000 distinct samples. They were simulated through the axonal injury model to generate outputs for CN training and prediction.
The CNN was trained to simultaneously predict the four axonal injury responses from the distinct FS dataset. Model training was iteratively tuned and evaluated with five-fold cross-validation until the desired performance was achieved. The final model was further validated using distinct FS profiles from the reserved for independent testing. Once trained, the CNN enabled rapid inference (< seconds) of multimodal axonal responses across the entire white matter by using the FS profiles of peak magnitude for each voxel, chosen to be 2 mm in isotropic resolution (92,523 voxels).

Results, Conclusions, and Discussions::

Results:
The CNN achieved high accuracy in prediction for all four axonal response. With 2000 FS samples, five-fold cross-validation yielded an average R² of 0.935 and scaled RMSE of 0.0095. In independent testing using 75 FS profiles from the reserved head impact, all R² values exceeded 0.91. The RMSE was lowest for tau failure (~4.0%) and highest for NF failure (~7.5%), confirming the model’s robust generalization to unseen data.

After validation, the CNN was applied to generate voxelwise predictions for the concussive impact across the entire white matter. The model produced spatially detailed maps of MT peak MPS, axolemma peak MPS, tau failure percentage, and NF failure percentage (Figure 1). No spatial correlations of “hot spots” appear to exist among these

The full-brain inference completed in ~12 seconds on a standard workstation. This contrasted sharply against conventional FE simulation, which was estimated to require ~12 years. These results highlight the model’s unique capability to support large-scale or time-sensitive applications in future TAI studies.

Discussion:
This work demonstrates that a CNN trained with a stratified and adaptive FS dataset can accurately and efficiently estimate microscale axonal responses across the white matter. By using diverse FS profiles from a stratified and adaptive sampling strategy, the model reduced the need for costly simulations for generating the training samples while preserving generalizability. Testing confirmed robust performance, and the CNN enabled rapid generation of voxelwise injury maps to capture multimodal axonal responses of different substructures. These capabilities enable seamless integration of biomechanics with imaging biomarkers, personalizing TBI risk assessments, or conducting population-level studies. This framework could significantly advance multiscale modeling by linking tissue deformation to localized axonal responses with near real-time inference.

Conclusions:
The trained CNN enables fast, accurate prediction of whole-brain axonal injury, bridging computational efficiency and multiscale insight for TBI research and application.

Acknowledgements and/or References (Optional)::

References:
Zhang C., Ji S., Sex differences in axonal dynamic responses under realistic tension using finite element models, J Neurotrauma 40 (2023) 2217–2232. https://doi.org/10.1089/NEU.2022.0512.
Zhang C., Ji S., Micromechanics of axonal injury in rapid tension and compression, Stapp Car Crash J 68 (2024). https://stapp.org/wp-content/uploads/SC24-02.pdf.