Deep TMS Targeting via SimNIBS + Unsupervised Analysis

Tags: Computational Neuroscience · Biomedical Engineering · SimNIBS · TMS · Finite Element Simulation · Python · Clustering

This project investigates whether deep transcranial magnetic stimulation (TMS) coils can effectively target deep brain structures commonly addressed by deep brain stimulation (DBS), using biophysical simulation and data-driven analysis.

Using SimNIBS, I ran scalp-wide coil placement optimizations across multiple head reconstructions and then analyzed the resulting optimal coil coordinates in MNI space using clustering and visualization techniques. A second-stage simulation validated field intensity at the target regions using the cluster-derived optimal coordinates.

High-Level Pipeline

Head model inputs: subject MRI-derived reconstructions and meshes (T1/T2 → headreco output).
Optimization runs: simulate coil placements across the scalp to find the best location for each target.
Coordinate normalization: convert subject-space outputs into MNI space via affine transforms.
Unsupervised analysis: identify consistent “best” scalp regions using clustering + t-SNE.
FEA validation: re-simulate using representative optimal coordinates to quantify induced field intensity.

Targets

Simulations were run for deep structures motivated by DBS targets / clinical relevance, including:

Amygdala
Anterior Cingulate Cortex
Fornix
Globus Pallidus
Hippocampus
Periaqueductal Gray
Subthalamic Nucleus
Ventromedial Thalamic Nucleus

What’s in the Repo

Optimization scripts: per-structure Python scripts (e.g., find_optimal_coords_for_*.py) that run SimNIBS optimization to estimate best scalp coil placement for each target.
Coordinate transform utilities: interpret_affine_coordinates.py for mapping between subject coordinates and MNI coordinates.
Analysis: kmeans_clustering.py, agglomerative_clustering.py, and TSNE_visualization.py to identify trends and stable placement clusters across subjects.
FEA follow-ups: FEA_simulations_with_optimal_coords*.py to quantify induced field intensity at the targets using cluster-derived coordinates.
Example data + outputs: representative head mesh inputs and sample SimNIBS optimization outputs, plus an example .msh file containing an e-field intensity map viewable in Gmsh/SimNIBS.

Why It Matters

Deep TMS is non-invasive and more accessible than DBS, but depth targeting is technically challenging.
This project uses simulation to systematically test whether deep coils can meaningfully reach subcortical targets.
The analysis identifies robust scalp placement regions and supports follow-up validation and future studies.

Disclaimer