Comprehensive evaluation of U-Net based transcranial magnetic stimulation electric field estimations

Transcranial Magnetic Stimulation (TMS) is a non-invasive method to modulate neural activity by inducing an electric field in the human brain. Computational models are an important tool for informing TMS targeting and dosing. State-of-the-art modeling techniques use numerical methods, such as the finite element method (FEM), to produce highly accurate simulation results. However, these methods operate at a high computational cost, limiting real-time integration and high throughput applications. Deep learning (DL) methods, particularly U-Nets, are being investigated for TMS electric field estimations. However, their performance across large datasets and whole-head stimulation conditions has not been systematically evaluated. Here, we develop a DL framework to estimate TMS-induced electric fields directly from an anatomical magnetic resonance image (MRI) and TMS coil parameters. We perform a comprehensive evaluation of the performance of our U-Net approach compared to the FEM gold standard. We selected a dataset of 100 MRI scans from a diverse population demographic (ethnic, gender, age) made available by the Human Connectome Project. For each MRI, we generated a FEM head model and simulated the electric fields for 13 TMS coil orientations and 1206 positions (a total of 15,678 coil configurations per participant). We trained a modified U-Net architecture to predict individual TMS-induced electric fields in the brain based on an input T1-weighted MRI scan and stimulation parameters. We characterized the model's performance according to computational efficiency and simulation accuracy compared to FEM using an independent testing dataset. The U-Net results demonstrated an accelerated electric field modeling speed at 0.8 s per simulation (×97,000 times acceleration over the FEM-based approach). Sampling stimulation conditions across the whole brain yielded an average DICE coefficient of 0.71 ± 0.06 mm and an average center of gravity deviation of 7.52 ± 4.06 mm from the FEM-based approach. Our findings indicate that while deep learning has the potential to significantly accelerate electric field predictions, the precision it achieves needs to be evaluated for the specific TMS application.