Abstract
Transcranial magnetic stimulation (TMS) is often applied to the motor cortex to stimulate a collection of motor evoked potentials (MEPs) in groups of peripheral muscles. The causal interface between TMS and MEP is the selective activation of neurons in the motor cortex; moving around the TMS ‘spot’ over the motor cortex causes different MEP responses. A question of interest is whether a collection of MEP responses can be used to identify the stimulated locations on the cortex, which could potentially be used to then place the TMS coil to produce chosen sets of MEPs. In this work we leverage a 3D convolutional neural network (CNN) architecture to tackle an inverse imaging task in which we start with the MEPs and estimate the stimulated regions on the motor cortex. We present and evaluate five different inverse imaging CNN architectures, both conventional and generative, in terms of several measures of reconstruction accuracy. We found that one architecture, which we propose as M2M-InvNet, consistently achieved the best performance.
Introduction
Transcranial Magnetic Stimulation (TMS) is a non-invasive tool used to stimulate specific brain regions. When applied to the motor cortex, it triggers "Motor Evoked Potentials" (MEPs)—electrical signals in peripheral muscles. While researchers have successfully addressed the "forward problem" (predicting muscle responses from brain stimulation), the inverse problem remains a significant challenge: Given a specific pattern of muscle twitches, exactly which part of the brain was stimulated and how intense was the field?
Solving this inverse problem would allow clinicians to precisely place TMS coils to achieve desired therapeutic effects or to map changes in brain-to-muscle "wiring" following neurological events like a stroke. This task is difficult because the mapping is complex, noisy, and "many-to-one," where different stimulation patterns might produce similar muscle responses.
Methods
The study evaluated five different 3D Convolutional Neural Network (CNN) architectures for this reconstruction task, categorized into two main groups:
- Conventional CNNs: Standard architectures like Autoencoders (AE) that attempt to map inputs to outputs directly.
- Variational/Generative Models: Architectures utilizing Variational Inference (VI), such as Variational Autoencoders (VAEs), designed to ensure reconstructed E-fields remain on a "biologically meaningful manifold."
The proposed architecture, M2M-InvNet (Direct Variational), employs a single-stage training strategy. Rather than learning brain E-field patterns and muscle mapping separately, it jointly learns the latent relationship between the two. This approach better captures the nuances of how specific field shapes translate to multi-muscle activations.
Datasets and Ground Truth
Data was collected from three healthy right-handed subjects, with approximately 1,000 scalp locations stimulated per subject across four intensities (110% to 140% of the resting motor threshold). Since direct measurement of electric fields inside a living brain is impossible, high-resolution T1 MRI scans were used to create subject-specific head models. Finite Element (FE) modeling then simulated the "ground truth" E-fields for every stimulation. Electromyography (EMG) recorded MEPs from 15 different hand and arm muscles as the primary response data.
Key Results
M2M-InvNet achieved the highest performance across all subjects, with a Normalized Root Mean Square Error (NRMSE) as low as 0.156 and an $R^2$ value up to 0.855. Qualitatively, it was the only model that accurately reproduced both the intensity profile and the Center of Gravity (CoG) of the induced electric fields. Other models tended to underestimate peak stimulation, effectively "flattening" the intensity.
Generative constraints proved crucial; models using Variational Inference consistently outperformed purely convolutional ones. By constraining the latent space, the models were forced to learn the underlying structure of brain stimulation rather than memorizing noise. Furthermore, the model's accuracy is dependent on input quality; it performs best when there is a robust, multi-muscle response pattern to analyze.
Limitations and Future Work
The current study is limited by a small sample size of three healthy subjects. Future work must expand this to larger, more diverse populations, including clinical patients. Additionally, while FE modeling is the gold standard for ground truth, it remains a simulation with inherent assumptions. The authors suggest integrating Active Learning to optimize stimulation locations during sessions and aim to use deep learning to eventually replace the computationally expensive FE modeling pipeline entirely.
Summary updated