Machine learning based prediction of single-frequency viscoelastic brain white matter - A data science framework.

Area of Science:

• Neuroscience
• Biophysics
• Computational Biology

Background:

• In vivo Magnetic Resonance Elastography (MRE) and Diffusion Tensor Imaging (DTI) are costly and time-consuming for brain white matter (BWM) characterization.
• Existing numerical modeling approaches like finite element models (FEMs) have limitations in accuracy and computational resources for complex brain tissue behavior.

Purpose of the Study:

• To develop a cost-effective machine learning (ML) workflow for predicting the homogenized viscoelastic properties of BWM.
• To create a surrogate model using FEM-derived data to overcome the scarcity of experimental data.

Main Methods:

• A triphasic 2D composite model simulating BWM under shear stress was used to generate a synthetic FEM dataset.
• Machine learning regression models were trained using microstructural features (fiber volume fraction, moduli, axonal geometry).
• Feature selection and hyperparameter optimization were employed to enhance prediction accuracy, with decision tree-based models showing superior performance.

Main Results:

• Machine learning models successfully predicted the frequency-dependent mechanical response of BWM.
• SHAP interpretation identified glial moduli and fiber volume fraction as key predictors of mechanical properties.
• Decision tree-based ML models demonstrated high accuracy in predicting viscoelastic properties.

Conclusions:

• The proposed ML framework provides a computationally efficient and cost-effective alternative to in vivo MRE/DTI and direct FEM simulations for BWM characterization.
• This approach lays the groundwork for future ML-driven inverse models to link brain tissue constituents to neuroimaging characteristics.
• Potential applications include informing studies on neurodegenerative diseases like aging, dementia, and traumatic brain injuries.