Efficient TMS-Based Motor Cortex Mapping Using Gaussian Process Active Learning

Razieh Faghihpirayesh, Mathew Yarossi, Tales Imbiriba, Dana H. Brooks, Eugene Tunik, Deniz ErdoğmuşIEEE Transactions on Neural Systems and Rehabilitation Engineering 2021 · DOI

TL;DR

This paper introduces novel machine learning techniques, specifically Gaussian Process (GP) active learning, to automate and significantly improve the efficiency of Transcranial Magnetic Stimulation (TMS) for mapping the motor cortex. By intelligently selecting stimulation sites, these methods achieve comparable or better accuracy than human-guided approaches in a fraction of the time, paving the way for fully automated TMS mapping.

What Problem Does This Paper Address?

Mapping the motor cortex using Transcranial Magnetic Stimulation (TMS) is crucial for understanding brain function, surgical planning, and studying neurological conditions. The standard method involves stimulating many points on the scalp and recording Motor Evoked Potentials (MEPs). However, this process is notoriously time-consuming and inefficient. This is because responsive areas are sparse and spatially correlated, meaning many stimulated points yield no useful data. Traditional approaches like grid sampling are slow, and even expert-guided methods (User-Guided, USRG) rely on subjective operator experience. The challenge is to develop an automated, efficient, and accurate method for TMS mapping that can intelligently explore the scalp and focus on informative regions.

Methods

The core of this work lies in applying Gaussian Process (GP) active learning to guide TMS stimulation. GPs are powerful probabilistic models that can learn a function from data and, importantly, provide an estimate of uncertainty about their predictions. This uncertainty is key for active learning.

The paper explores several GP-based strategies:

  • Standard GP Entropy (GPE): Selects the next stimulation site that maximizes the model's uncertainty (entropy) about the MEP response.
  • Amplitude-Based GP Entropy (GPEμ): An improvement on GPE that considers both uncertainty and the amplitude of MEPs, aiming to better mimic human expert behavior.
  • Warped GP Entropy (WGPE): Addresses the non-negativity of MEPs by using a "warping" function, allowing GPs to model these inherently positive values more accurately. This method also balances exploration and exploitation effectively.
  • Adaptive GP Random Sampling (GPRS): A non-grid-based approach that uses a learned probability distribution to guide random sampling, concentrating on areas of interest.

These computational methods are compared against traditional approaches like regular grid sampling (GRID), random sampling (RAND), and the user-guided (USRG) method.

Datasets

The study utilized data collected from five healthy human subjects. For computational experiments, a "ground truth" (GT) map was created by combining and interpolating the experimental maps from GRID, RAND, and USRG. This GT served as a benchmark for evaluating the accuracy of the proposed computational methods. The paper also mentions using a 2mm resolution grid for selectable loci for the active GP methods, and a 0.5mm resolution grid for testing data.

Key Insights and Ideas

  • Automated Expertise: The research demonstrates that active learning with GPs can effectively replace human operator expertise in guiding TMS stimulation, leading to more objective and repeatable results.
  • Exploration vs. Exploitation Balance: The key to efficient mapping is balancing exploring uncertain areas with exploiting known responsive regions. WGPE and GPRS excel at this by leveraging GP uncertainty and learned distributions, respectively.
  • Non-negativity Matters: MEPs are inherently non-negative. Using Warped GPs (WGPE) to account for this biological constraint significantly improves mapping accuracy.
  • Beyond Grids: The GPRS method shows that efficient mapping doesn't necessarily require a predefined grid, reducing computational cost and offering more flexibility.

Key Results

  • Efficiency Gains: The proposed computational methods, particularly WGPE and GPRS, achieved comparable or better accuracy (lower Normalized Mean Squared Error - NMSE) than the human-guided USRG method, often with fewer stimuli.
  • USRG Superiority: The USRG method was significantly better than GRID and RAND, highlighting the value of intelligent, feedback-driven sampling.
  • Top Performers: WGPE and GPRS achieved the lowest NMSE, with WGPE even outperforming USRG in some cases. For example, WGPE and GPRS achieved lower GM NMSE than USRG after 100 simulations.
  • Time Savings: The paper suggests that these automated methods can drastically reduce mapping time, potentially from an hour to much less, by efficiently focusing on relevant areas.

Limitations and Future Work

  • Canonical Brain: The study used a canonical brain model for neuronavigation, not individual MRI reconstructions. This might have impacted the USRG operator's ability to use subject-specific anatomy.
  • Experienced Operator: Experiments were conducted by an experienced TMS operator. Future work should investigate inter-operator reliability and the performance with novice operators.
  • Single Muscle Mapping: Only one muscle (FDI) was mapped. Future research should explore multi-muscle mapping and how user-guided approaches might be improved with feedback from multiple muscles.
  • Coil Orientation and Intensity: The current algorithms do not dynamically adjust coil orientation or intensity. Future work aims to incorporate these factors and individual head models for even more precise mapping.

Summary updated

Abstract

Transcranial Magnetic Stimulation (TMS) can be used to map cortical motor topography by spatially sampling the sensorimotor cortex while recording Motor Evoked Potentials (MEP) with surface electromyography (EMG). Traditional sampling strategies are time-consuming and inefficient, as they ignore the fact that responsive sites are typically sparse and highly spatially correlated. An alternative approach, commonly employed when TMS mapping is used for presurgical planning, is to leverage the expertise of the coil operator to use MEPs elicited by previous stimuli as feedback to decide which loci to stimulate next. In this paper, we propose to automatically infer optimal future stimulus loci using active learning Gaussian Process-based sampling in place of user expertise. We first compare the user-guided (USRG) method to the traditional grid selection method and randomized sampling to verify that the USRG approach has superior performance. We then compare several novel active Gaussian Process (GP) strategies with the USRG approach. Experimental results using real data show that, as expected, the USRG method is superior to the grid and random approach in both time efficiency and MEP map accuracy. We also found that an active warped GP entropy and a GP random-based strategy performed equally as well as, or even better than, the USRG method. These methods were completely automatic, and succeeded in efficiently sampling the regions in which the MEP response variations are largely confined. This work provides the foundation for highly efficient, fully automatized TMS mapping, especially when considered in the context of advances in robotic coil operation.
BibTeX
@article{faghihpirayesh2021efficient,
  author = {Razieh Faghihpirayesh and Mathew Yarossi and Tales Imbiriba and Dana H. Brooks and Eugene Tunik and Deniz Erdoğmuş},
  title = {{Efficient TMS-Based Motor Cortex Mapping Using Gaussian Process Active Learning}},
  journal = {IEEE Transactions on Neural Systems and Rehabilitation Engineering},
  year = {2021},
  doi = {10.1109/tnsre.2021.3105644},
}
Figure