Rapid whole-brain electric field mapping in transcranial magnetic stimulation using deep learning

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Transcranial magnetic stimulation (TMS) is vital for treating neuropsychiatric disorders and neuroscience research.
• Accurate, subject-specific brain region targeting is crucial for TMS efficacy.
• Current numerical methods for estimating TMS-induced electric fields (E-field) are computationally intensive.

Purpose of the Study:

• To develop a deep neural network (3D-MSResUnet) for rapid, whole-brain E-field estimation in TMS.
• To accelerate the prediction of E-field distribution for personalized TMS applications.
• To reduce the computational time for TMS targeting.

Main Methods:

• A novel deep neural network, 3D-MSResUnet, integrating 3D U-net, residual modules, and multi-scale feature fusion.
• Training the network using a large dataset of finite element method (FEM) computed E-fields and diffusion MRI-derived conductivity or anatomical images.
• Evaluating the network's performance using various metrics, imaging modalities, and coil configurations.

Main Results:

• The 3D-MSResUnet accurately estimates TMS-induced E-fields, comparable to state-of-the-art FEM methods.
• Prediction time was drastically reduced to approximately 0.24 seconds per subject.
• The network demonstrated reliable performance across different imaging inputs and coil types.

Conclusions:

• Deep neural networks, specifically 3D-MSResUnet, offer a powerful and efficient tool for accelerating E-field prediction in TMS.
• This approach has the potential to significantly enhance the precision and accessibility of TMS targeting.
• AI-driven E-field estimation can advance personalized neurostimulation therapies.