A generalized workflow for conducting electric field-optimized, fMRI-guided, transcranial magnetic stimulation

Abstract

Transcranial magnetic stimulation (TMS) is a noninvasive method to stimulate the cerebral cortex that has applications in psychiatry, such as in the treatment of depression and anxiety. Although many TMS targeting methods that use figure-8 coils exist, many do not account for individual differences in anatomy or are not generalizable across target sites. This protocol combines functional magnetic resonance imaging (fMRI) and iterative electric-field (E-field) modeling in a generalized approach to subject-specific TMS targeting that is capable of optimizing the stimulation site and TMS coil orientation. To apply this protocol, the user should (i) operationally define a region of interest (ROI), (ii) generate the head model from the structural MRI data, (iii) preprocess the functional MRI data, (iv) identify the single-subject stimulation site within the ROI, and (iv) conduct E-field modeling to identify the optimal coil orientation. In comparison with standard targeting methods, this approach demonstrates (i) reduced variability in the stimulation site across subjects, (ii) reduced scalp-to-cortical-target distance, and (iii) reduced variability in optimal coil orientation. Execution of this protocol requires intermediate-level skills in structural and functional MRI processing. This protocol takes ~24 h to complete and demonstrates how constrained fMRI targeting combined with iterative E-field modeling can be used as a general method to optimize both the TMS coil site and its orientation.

Introduction

Transcranial magnetic stimulation (TMS) is a noninvasive method to stimulate the cerebral cortex by using powerful magnetic pulses applied to the scalp^1,2. With the advent of neuronavigation and figure-8 coils, it is possible to target specific cortical regions with sub-centimeter accuracy^3-5. Capitalizing on this improved spatial resolution, it is possible to use TMS to modulate ongoing cognitive processes with a high degree of functional specificity^6-10. However, to fully do so one must (i) possess prior knowledge of the cortical localization of the function of interest and (ii) use this knowledge to operationally define subject-specific targets.

Comparison with other methods

Although there are many approaches to TMS targeting^11-13, including the 5-cm rule commonly used to target the dorsolateral prefrontal cortex (dlPFC) in depression protocols^14-16, greater precision can be achieved by defining subject-specific targets with fMRI and using neuronavigation^17-19. fMRI, a widely used method for studying brain activity during cognitive paradigms, can deliver subject-specific, high-resolution maps of the entire brain that reflect changes in metabolic activity due to task demands^20-22. Numerous cognitive paradigms have been adapted to the MR environment, making it possible to characterize the neural activity mediating a diverse set of cognitive processes^23,24. Sack et. al.¹⁹ showed that targeted TMS using single-subject fMRI-guided TMS yields the largest effect size as compared with electroencephalogram (EEG) 10-20 system-guided, single-subject structural MRI-guided, and group fMRI-guided TMS. However, despite the obvious benefits of fMRI targeting, localization of function with fMRI data can be variable when using single-subject data, especially in frontal regions, where the anatomy can vary greatly across subjects^25-27. In addition, because the signal-to-noise ratio is often low with fMRI, it is unclear how much of the intersubject variability in fMRI-guided TMS target location is due to real anatomical differences as opposed to noise in the fMRI data. Therefore, when using fMRI to guide TMS targeting, it is critical to constrain the variability in potential targets by using prior knowledge of the underlying anatomy.

In addition to high variability in functional localization across subjects, fMRI is capable of informing only three of the six possible movement axes in a figure-8 TMS coil: those that determine the x, y, z positions. The effects of TMS on neural function depend in part on the orientation of the coil with respect to the anatomical ROI^28-31; therefore, it is important to also consider the other three movement axes (roll, pitch, and yaw) that determine coil orientation when selecting a stimulation site. Recent advances in E-field modeling have made it possible to estimate the amplitude and directionality of current induced in the brain, given a coil location and orientation^32-37. Accordingly, by iterating E-field models across multiple orientations, it is possible to assess the potential efficacy of each model and choose the optimal orientation.

Development of the protocol

At a fundamental level, this protocol uses fMRI to choose optimal TMS sites for individual subjects and E-field modeling to optimize coil orientation at the stimulation site. The protocol was used in previous task-based and resting-state fMRI studies to target frontal and parietal regions implicated in anxiety^38,39. The first study targeted the right dlPFC³⁸. The rationale for this targeting approach was to enhance working memory (WM) processes in the right dlPFC using high-frequency (10-Hz) repetitive TMS (rTMS) and measure the effect of this manipulation on anxiety-potentiated startle (APS) using threat of unpredictable shock. We used previous fMRI data collected during a Sternberg WM task to create a binary mask of the right dlPFC. We then used the Sternberg WM paradigm to drive blood oxygenation level-dependent (BOLD) activity in the right dlPFC at the single-subject level. By starting with a functional mask, we were able to constrain the location of the fMRI peaks obtained at the single-subject level to a predefined right dlPFC region that had previously been shown to be important for WM-related processing. The second study targeted the intraparietal cortex (IPS)³⁹ based on previous work suggesting that this region was a connectivity hub for anxiety related processes⁴⁰. The rationale for this targeting approach was to reduce excitability in this connectivity hub using low-frequency (1-Hz) rTMS. We drew 10-mm spheres around bilateral group-level peak coordinates in the IPS. We then collected resting-state fMRIs for each subject, computed global brain connectivity, and selected the voxel within these search spheres with the strongest connectivity to the rest of the brain. In both studies we were able to identify reliable individualized TMS targets using this combined group-level masking/single-subject mapping approach.

The E-field modeling approach used in this protocol was developed by Thielscher et al.³⁷, who developed the SimNIBS software package. In the studies mentioned above, we build on the work by Thielscher et al. by iterating the creation of the E-field model by varying yaw vector orientations and sampling the resulting maps using the group-level ROIs described above^38,39. This approach enabled us to summarize and extract useful data from the E-field models and apply these in a systematic way across subjects to inform coil placement. Key to this development is the underlying assumption that maximizing the induced current in the theoretical ROI (i.e., group-level masks of the right dlPFC or IPS) would maximize the potential effect of the TMS stimulation. This assumption should be rigorously tested in future work.

During the development of this protocol we were guided by the desire to create a simple and reliable data analytic strategy that could optimize all positional axes needed to define coil placement for each subject. One of the challenges we faced during development was that different software platforms (i.e., AFNI, FSL, FreeSurfer, SimNIBS, Brainsight) have different conventions for reading/writing data. These differences include variations in axis orientation coding (i.e., right anterior superior (RAS) versus right anterior inferior (RAI) versus left posterior inferior (LPI)), different coordinate systems (i.e., RAS versus surface-RAS versus native space versus Montreal Neurological Institute (MNI)), different geometries (e.g., volume-based versus surface-based), and different file types (e.g., MGH-NMR gzipped versus NIfTI). These differences can lead to misregistrations across files created by different packages. Although large errors are often easily spotted, small errors can go unnoticed. In addition, it is often not clear how to transform data (especially target coordinates) to correct these errors. To overcome these challenges, we implemented the following basic strategies. (i) Generate the head model using SimNIBS before registering the echo planar images (EPIs) and the group-level ROI mask to the T1 images. This principle ensures that the volume-based EPI and mask data are properly registered to the surfaced-based head model. (ii) Store all necessary targeting data as volumetric NIfTI files that are co-registered to the subject’s surface-aligned T1 image. For instance, we would draw a 5-mm sphere centered on the target coordinates for a subject’s stimulation site and save that as a NIfTI file. This principle is to avoid misreading of coordinate files (i.e., interpreting a set of coordinates as LPI when they are encoded as RAI), because the T1 and the coordinate overlay are encoded identically. (iii) Finally, visualize the data before and after any spatial transformation. This principle is to ensure that the transformation was applied properly.

Advantages and limitations

Using fMRI to guide the site of stimulation is currently a commonly used method in rTMS targeting for neuroscience studies¹⁹. Here, we describe a method that uses standard MRI and fMRI scans and freely available software to conduct fMRI-guided rTMS. This method improves upon previous methods because it accounts for the highly variable spatial patterns in single-subject fMRI. Furthermore, it uses iterative E-field modeling to optimize the TMS coil handle orientation. Importantly, the approach described here is not task or region specific. Instead, it is generalizable to any clearly defined theoretical target region. One of the advantages of our protocol is that it is agnostic to target region specification, meaning that if users aims to target a small patch of cortex, they can sample the E-field using a small spherical region at the area of the individual fMRI peak activation. Alternatively, if users aim to target an entire region, they can sample the E-field with a larger regional ROI. This enables users to optimize their targeting using their a priori assumptions about the cortical involvement in the behavior that they are trying to modify; however, ROI size selection should be based on prior data demonstrating the extent of the anatomical/functional region targeted.

One limitation of the current protocol is that it does not directly account for the relative angle between the cortical surface and the local E-field. Instead, the protocol uses the magnitude of the E-field, irrespective of the current direction. The extent to which this plays a role in neural activation from TMS is an open question. For example, Weise et al.⁴¹ showed that the field tangent to the cortex at M1 is critical for determining motor evoked potential (MEP) amplitude. By contrast, Aberra et al.⁴² reported that magnitude, irrespective of direction, is an important predictor of neural activation. Accordingly, it would be useful to include an option to optimize the E-field in terms of vector component relative to specific gyrus or sulcus within the ROI, which can be done during the E-field modeling steps by selecting the E-field vector output in SimNIBS. However, we have not done this because it would be difficult to noninvasively test the predictions of these two models (E-field vector component versus magnitude) outside of the motor cortex, where a reliable behavioral measure of neural activation does not typically exist. In spite of this limitation, the protocol described here is agnostic with respect to which of these two metrics is used to quantify the modeled TMS effect, and it would be rather straightforward for a user to replace the E-field norm map with an E-field map weighted by the relative angle between the vector direction and the local norm of the cortical surface. Another limitation is that it requires between 24 and 48 h between data collection and any TMS delivery, making single-session scanning/TMS impossible.

Experimental design

In this protocol, we describe a generalized approach to optimize the site of stimulation and coil orientation for TMS (i.e., for all six movement axes) using fMRI and E-field modeling (see Box 1 and Fig. 1 for an overview). To apply this approach, users will need three key elements: (i) high-resolution structural and diffusion scans, (ii) EPI scans capable of generating reliable single-subject fMRI data, and (iii) a structural mask of the target region that can constrain the search for single-subject targets.

Box 1 ∣. Definition of parameters used to optimize coil position in TMS.

The purpose of this protocol is to provide a generalized approach to optimize coil position during TMS. Here, we define coil position in terms of location in Euclidean space (i.e., x, y, z coordinates) and orientation (roll, pitch, and yaw). Location is determined using single-subject functional neuroimaging data (e.g., task-based or resting state fMRI). Roll and pitch are perpendicular to the scalp at the stimulation site. Yaw is determined using electric-field modeling.

Fig. 1 ∣. Schematic of the steps described in the Procedure.

a, Example of a group-level target mask from functional data collected during the Sternberg WM task. b, BOLD data from a single subject collected during the Sternberg WM task. c, Surface mesh created by the mri2mesh algorithm. d, Peak activation within target mask from BOLD image in b. e, E-field simulation targeting coordinates depicted in d. f, Polar plot for a single subject representing E-field magnitude as a function of coil orientation visualized on the scalp for use in neuronavigation software. g, Mark for TMS coil placement from a session targeting coordinates in d. Labels on figure correspond to section headings in protocol.

To apply this protocol, users must first decide on a cortical target and operationally define this target by creating an ROI mask in standard space. This decision should be based on previous research. It is critical that users can reliably localize their ROI based on single-subject fMRI data. Next, users should collect the necessary MRI/fMRI data and process the data at the single-subject level. Structural data should be used to create a realistic head model for E-field calculations. Functional data will be used to identify the optimized site of stimulation at the single-subject level. Using the head model and the coordinates for the subject-specific stimulation site, users should then conduct a series of E-field models, varying the angle of the yaw vector for each simulation. Finally, users should evaluate the E-field models to determine which yaw vector induces the largest amount of current in the target region.

To develop this protocol, we targeted the right dlPFC and operationally defined this region using three different a priori target definitions (anatomical, functional, and meta-analytical) and a fourth target definition (unconstrained), which is not recommended but is included to show the need to constrain the target search using a priori anatomical definitions when using single-subject fMRI data for targeting. This target definition step should precede data collection because the data should be analyzed at the single-subject level in real time to inform TMS targeting on a subject-by-subject basis. We found little difference in target estimates when using the functional mask versus the meta-analytical mask derived from the NeuroSynth database (https://neurosynth.org/). Accordingly, NeuroSynth may be a convenient method for operationally defining a theoretical target when existing functional data are not available.

To identify the optimal right dlPFC target at the single-subject level, we administered the Sternberg WM task^43-45 (Fig. 2) to subjects while recording BOLD activity. In each trial, subjects were presented with a series of letters and were instructed to either remember the letters in the order they were presented (maintain) or rearrange the letters in alphabetical order (sort). At the response prompt, participants were presented with a letter and a number and had to indicate whether the position of the letter in the series matched the number. There were 52 total trials with three trial types: maintain-five-letter trials, maintain-eight-letter trials, and sort-five-letter trials. We used the contrast between the maintain-five-letter trials and the sort-five-letter trials to identify activity related to WM manipulation (Fig. 2). We calculated the difference between BOLD activity during the sort-five trials and that during the maintain-five trials as a measure of WM manipulation related BOLD activity. We chose this task because it is known to reliably activate the dlPFC^43-45, which is important because single-subject fMRI contrasts have very low signal-to-noise ratios^25-27. Users of this protocol should identify a similarly robust task with which to study their target ROI. In addition to choosing a robust task to activate the right dlPFC, we used our a priori target ROI masks to limit the search region for individual fMRI peaks, meaning that the stimulation site was defined as the voxel within the ROI with the peak activation difference for our sort-versus-maintain contrast. This step is critical to overcoming the inherent noisiness of the single-subject fMRI data.

Fig. 2 ∣. Schematic demonstrating the use of the task throughout the targeting and stimulation steps.

a, We recommend using a task (Sternberg WM paradigm in our case) to derive a functional ROI centered on the target region (see ‘Target definition’ section of the Procedure). Each trial of the Sternberg task consists of encoding, maintenance, and retrieval periods separated by a variable intertrial interval (ITI). During the encoding period, subjects are presented with a series of letters. During the maintenance period, the subjects are required to retain the series of letters in WM. During the retrieval period, subjects indicate whether the position of the letter in the series matches the number. b, We recommend using the same task to generate the BOLD data to guide the TMS targeting (see ‘Data collection’ section of the Procedure). c, We recommend using the same task during stimulation to capitalize on paired associative stimulation processes (see ‘Neuronavigation’ section of the Procedure). During online stimulation, the TMS train can be delivered during the maintenance interval.

After identifying the stimulation site for each subject, we then conducted E-field models, to estimate the distribution of current likely to occur if we stimulated our site of interest. Given that orientation of the coil relative to the cortical anatomy is known to affect current distribution, we conducted multiple E-field models, varying the yaw vector orientation at 15-degree increments around a circle perpendicular to the scalp at the site of stimulation. We then evaluated these E-field models against one another by estimating the overall current distribution at either the site of stimulation (defined by a 5-mm sphere) or across the entire ROI (defined by the ROI mask used in previous steps). Across subjects, estimates of the current distribution were more variable when sampled at the site of stimulation versus across the entire ROI.

On the basis of the results presented here, we recommend (i) generating a functional mask from task-based BOLD activity in an independent group of subjects, (ii) creating single-subject targets within this mask from task-based BOLD activity, (iii) optimizing the TMS coil handle vector with iterative E-field modeling evaluated using ROI-level statistics, and (iv) using the task to activate the target region during the task to capitalize on paired-associative stimulation effects. Following these recommendations is expected to result in several improvements consistent with more accurate targeting, such as more consistent targets across subjects, more reliable E-field model estimates, and more consistent coil handle vectors across subjects. Given that TMS dosing is affected by details such as scalp-to-cortex distance⁴⁶, these improvements should also result in more consistent dosing across subjects.

Materials

Human participants

• Volunteers must not have any MRI or TMS contraindications. Participants were recruited from the community via flyers, advertisements, and listservs ! CAUTION Written informed consent must be obtained from participants. All subjects gave written informed consent and were compensated for their time; our study was approved by the National Institute of Mental Health (NIMH) Combined Neuroscience Institutional Review Board.

Equipment

Starting data

• This protocol requires collection of T1-, T2- and diffusion-weighted images (DWIs). In addition, it requires the collection of either task-based or resting-state fMRI data

• The parameters used to obtain scans used in the example procedure are summarized in Table 1. These images were collected on a Siemens 3T Skyra MRI scanner with a 32-channel head coil

• T1- and T2-weighted images should be high resolution (<1 mm isotropic) and cover the whole brain and scalp. The inclusion of T2-weighted images is optional but is recommended to improve tissue segmentation accuracy

• Diffusion-weighted images (DWIs) should be of adequate resolution (<2 mm isotropic), single shell, and have a minimum of 30 directions

• The parameters of the EPI images will depend on the task used for targeting, but the EPI data should be of sufficient resolution and quality to localize the target of interest

• An example dataset is included as Supplementary Data. A summary of the files included in the example dataset is provided in Box 2

Table 1 ∣.

Parameters used for MRI scans

Scans	Resolution (mm)	FOV	No. of slices	TR (ms)	TE (ms)	Flip angle (degrees)	Additional information
T1	0.8	256 × 256	176 sagittal	2,400	2.24	7
T2	0.8	300 × 320	208 sagittal	3,200	566	120
DWI	2	128 × 128	70 axial	12,000	94	90	B0 = 100; directions = 30. Single shell; no cardiac gating
EPI	3	64 × 64	32 axial	2,000	13.8, 31.2, and 48.6	70	Multi-echo sequence

All data were collected on a Siemens 3T Skyra scanner with a 32-channel head coil. B₀, gradient strength; FOV, field of view; TE, echo time; TR, repitition time.

Box 2 ∣. Files included in the example dataset.

The following sample data can be used to perform the analysis described in the Procedure.

• High-resolution T1 image: anat_T1w.nii (https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/01.mri2mesh.csh)

• High-resolution T1 image: anat_T2w.nii (https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/01.mri2mesh.csh)

• Diffusion-weighted image: dwi_dir-98_dir-ap.nii (https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/01.mri2mesh.csh)

  File containing B0 values for DWI run: dwi_dir-98_dir-ap.bval

  File containing B0 vectors for DWI run: dwi_dir-98_dir-ap.bvec

• Functional MRI time series: func_task-cog_run-01_bold_dir-ap.nii (https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/02.afni_proc.csh)

• Event timing files to accompany the functional MRI timeseries (https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/02.afni_proc.csh):

  • sub-000_task-cog-maintain-letter.1d
  • sub-000_task-cog-sternberg.button.1d
  • sub-000_task-cog-sort-letter.1d
  • sub-000_task-cog-sort-response.1d
  • sub-000_task-cog-sort-maintenance.1d
  • sub-000_task-cog-maintain-response.1d
  • sub-000_task-cog-maintain-maintenance.1d

Software

▴ CRITICAL This protocol is intended to be run on a modern workstation running an up-to-date Linux distribution (e.g., Fedora, Ubuntu) ▴ CRITICAL Scripts to execute this protocol on the sample data are provided on GitHub: https://github.com/balders2/tms_targeting ▴ CRITICAL Before conducting this protocol on the sample data, users should download both the data (Supplementary Data) and the scripts (https://github.com/balders2/tms_targeting) to their local machine. Next, users should download the software and dependencies listed below, using the instructions included with the distributions. Finally, users should execute the scripts in order, as described in the Procedure.

• AFNI (v.19.2.01 or greater: https://afni.nimh.nih.gov/node/4). This is a toolbox for analyzing and displaying fMRI data; it is used in this protocol as the main toolbox for processing 3D fMRI and E-field model images⁴⁷

• SimNIBS (v.2.1.2 or greater: https://simnibs.drcmr.dk/). This is a Python package for conducting E-field modeling for TMS and transcranial direct current stimulation (tDCS). SimNIBS also requires the following programs: Python (v.2.7:https://www.python.org/downloads/), FreeSurfer (https://surfer.nmr.mgh.harvard.edu/fswiki/DownloadAndInstall) (or MATLAB, for tissue segmentation; see below), and FSL (https://fsl.fmrib.ox.ac.uk/fsldownloads_registration/download/fsl=5,o_s=19,d_type=release/, for diffusion tensor data processing). SimNIBS depends on a coil definition file for the TMS coil used for stimulation. Although SimNIBS includes coil definition files for several common coils, it may be necessary to contact the device manufacturer for additional information regarding other coil types³⁷

• MATLAB (v.8.5 (R2015a) or greater: https://www.mathworks.com/downloads/). Several of the scripts referenced in this protocol are written in MATLAB and rely on external functions/toolboxes (i.e., Patch Normals and stlTools). These scripts have been uploaded to GitHub (https://github.com/balders2/tms_targeting). However, these processes could be implemented in other, open-source software environments (e.g., Python)

• Brainsight (v.2.3.5 or greater: https://www.rogue-research.com/2017/05/brainsight-2-3-5-released/). This is a program that tracks the position of the coil relative to the position of the head during TMS sessions. The files created in this protocol conform to the standard NIfTI file format (https://nifti.nimh.nih.gov/) and should therefore be compatible with most neuronavigation systems, such as Localite TMS Navigator (https://www.localite.de/en/products/tms-navigator/). However, the files created in this protocol have not been tested with other neuronavigation options

• For additional dependencies needed to execute the protocol on the included sample data, see https://github.com/balders2/tms_targeting

Procedure

▴ CRITICAL The Procedure does not require a specific directory structure or file-naming convention. However, users should modify the scripts to conform to their internal existing naming conventions and preprocessing pipelines. If the user’s data are Brain Imaging Data Structure (BIDS) compliant (https://bids.neuroimaging.io/), few modifications (e.g., modifying the scripts to recognize the user’s functional files) will be necessary.

Target definition ● Timing ~1 h

1. Before data collection, create a binary mask dataset of the targeted region using existing group-level fMRI contrast images, meta-analysis results from tools such as NeuroSynth⁴⁸, or anatomical regions from a preferred atlas. Although the location and extent of the mask will depend on the user’s theoretical target, the mask should be created as a set of 1s and 0s marking target versus nontarget voxels, respectively, in standard space (e.g., MNI space).

2. If the target mask is in a standardized coordinate space, save a copy of the template file in NIfTI format to the project directory to facilitate alignment of the mask to native subject space.

Data collection ● Timing ~2 h

• 3Collect a high-resolution (<1-mm isotropic) T1 and T2 scan to be used to segment the subject’s anatomy into the following compartments: skin, skull, CSF (cerebrospinal fluid), gray matter, and white matter.
• 4Collect a single diffusion-weighted scan of adequate resolution (<2 mm isotropic), single shell, and with a minimum of 30 directions that will be used to further inform the current spread in white matter.
• 5Collect the necessary task-based or resting-state EPI data needed to localize activity related to the function of interest. The functional data used for targeting should be chosen on the basis of the best available data for the targeted region.

Structural data processing ● Timing ~24 h

• 6
  Use the SimNIBS utility

  mri2mesh

  to segment the different tissue compartments and mesh the compartments for subsequent finite-element analysis (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/01.mri2mesh.csh). Execute the code by typing the following into a command-line terminal:

  mri2mesh --all ${subject} anat_T1w.nii anat_T2w.nii

• 7
  (Optional) Use the SimNIBS utility

  dwi2cond

  to convert the diffusion-weighted data into conductivity tensors (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/01.mri2mesh.csh). If diffusion data are not available, one can skip this optional step and assume isotropic tissue conductivity in Step 13.

  dwi2cond --all ${subject} dwi_dir-98_dir-ap.nii dwi_dir-98_dir-ap.
  bval dwi_dir-98_dir-ap.bvec

• 8
  Align the target region mask and the surface-aligned T1 (SimNIBS output;

  filename = ${subject}_T1fs_conform.nii.gz

  ) image created by SimNIBS. If the target region is obtained from group data in standard space, it will be necessary to align this mask to the subject’s native space, so that it is aligned well to the surfaces used in the E-field models (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/01.mri2mesh.csh).

  3dQwarp -allineate -blur 2 2 -iwarp \
  -base ${resdir}/MNI152_2009_template.nii.gz \
  -source ${subject}.anat.aparc_ss+orig \
  -prefix ${subject}.anat.aparc_ss
  3dNwarpApply \
  -nwarp "${subject}.anat.aparc_ss_WARPINV+tlrc" \
  -dxyz 2 \
  -master ${subject}.anat.aparc_ss+orig \
  -source ${resdir}/rdlpfc.fmask.nii \
  -prefix $subject.rdlpfc.fmask.nii

  ? TROUBLESHOOTING

Functional data processing ● Timing ~2-4 h

• 9
  Use the

  afni_proc.py

  script to conduct the standard fMRI preprocessing steps: slice time correction, despiking, volume registration, alignment to the T1, masking, blurring, scaling, and first-level time series regression with volume scrubbing (RMS (root mean square) >0.5 mm for task-based fMRI)⁴⁷: (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/02.afni_proc.csh).

  afni_proc.py
  -subj_id ${subject} \
  -script code/proc.${subject}.${functag}.csh \
  -scr_overwrite \
  -out_dir ${subject}.${functag}_proc \
  -copy_anat anat_T1w.nii \
  -dsets func_task-cog_run-01_bold_dir-ap.nii \
  -blocks tshift align tlrc volreg blur mask scale regress \
  -align_opts_aea \
  -giant_move \
  -cost lpc+ZZ \
  -tlrc_base MNI152_T1_2009c+tlrc \
  -tlrc_NL_warp \
  -volreg_align_to MIN_OUTLIER \
  -volreg_align_e2a \
  -mask_epi_anat yes \
  -blur_in_mask yes \
  -blur_size 2 \
  -regress_stim_times \
  sub-000_task-cog-maintain-letter.1d \
  sub-000_task-cog-sternberg.button.1d \
  sub-000_task-cog-sort-letter.1d \
  sub-000_task-cog-sort-response.1d \
  sub-000_task-cog-sort-maintenance.1d \
  sub-000_task-cog-maintain-response.1d \
  sub-000_task-cog-maintain-maintenance.1d \
  -regress_stim_labels ${stimlabels} \
  -regress_stim_types ${stimtypes} \
  -regress_basis_multi ${stimmodels} \
  -regress_motion_per_run \
  -regress_apply_mot_types demean deriv \
  -regress_censor_motion 0.3 \
  -regress_censor_outliers 0.15 \
  -regress_censor_first_trs 4 \
  -regress_apply_mask \
  -regress_run_clustsim no \
  -test_stim_files no \
  -remove_preproc_files

  ? TROUBLESHOOTING

• 10Identify the contrast of interest and extract the single-subject peak fMRI activation coordinates: (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/03.target.csh).

  3dcalc \
  -a ${subject}.${functag}_proc/stats.${subject}+orig"[${functag}-sort-maintenance#0_Tstat]" \
  -b ${subject}.${functag}_proc/stats.${subject}+orig"[${functag}-maintain-maintenance#0_Tstat]" \
  -c ${subject}.rdlpfc.fmask.nii \
  -expr "(a-b)*ispositive(c-.5)" \
  -prefix ${subject}.srt-mnt.rdlpfc.nii

• 11
  Finally, save these coordinates, both as a text file and as a spherical ROI (

  ${subject}.srt-mnt.rdlpfc.nii.gz

  ) that can be loaded into the neuronavigation software: (cortex target; see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/03.target.csh).

  3dclust \
  -quiet \
  -orient RAI \
  -1clip .01 3.5 2 \
  ${subject}.srt-mnt.rdlpfc.nii \
  ∣ head -n1 \
  ∣ tr -s ‘ ‘ \
  ∣ cut -d" " -f15-17 \
  ∣ tee ${subject}.srt-mnt.rdlpfc.1d
  3dUndump \
  -master ${subject}_T1fs_conform.nii.gz \
  -prefix ${subject}.srt-mnt.rdlpfc.sphere.nii \
  -srad 5 \
  -xyz \
  -orient RAI \
  ${subject}.srt-mnt.rdlpfc.1d

E-field modeling ● Timing ~20–30 min per model

• 12
  Create a template file for the E-field model calculations. To do so, first load the mesh file (

  ${subject}.msh

  ) created by mri2mesh (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/01.mri2mesh.csh) into the SimNIBS graphical user interface (GUI) and then check the surfaces for defects (see SimNIBS documentation)³⁷.
• 13Then load the definition file (included in the SimNIBS installation) that corresponds to your TMS coil and specify the ‘volume normalize’ option to use the conductivity tensors in the E-field model.
• 14Next, enter the fMRI activation coordinates as a new entry in the TMS position list.
• 15Check the option to output the normalized E-field map as a NIfTI file.
• 16In the SimNIBS GUI, save the settings as a MATLAB file, which will serve as a template for the E-field model calculations.
• 17Run the included https://github.com/balders2/tms_targeting/blob/master/mfiles/vect_target_wrapper.m MATLAB function to duplicate the template file and update the yaw vector for each orientation (default is 24 orientations at 15-degree increments) to be tested (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/04.efield_template.csh).

  matlab \
  -nosplash \
  -nojvm \
  -r "vect_target_wrapper('${expdir}','${subject}', ‘srt-mnt.rdlpfc’, ‘Medtronic_MCF_B65.ccd’)"

• 18Conduct E-field model by running the SimNIBS command-line utility once for each yaw vector to generate the E-field maps (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/05.simnibs.csh):

  foreach simulation ($sims)
  set ds=‘ls *mat ∣ head -n1’
  simnibs --cpus 1 ${ds}
  end

  ? TROUBLESHOOTING

Target selection ● Timing ~30 min

• 19Resample the E-field volumes to EPI resolution (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/06.3dresample.csh):

  foreach simulation ($sims)
  set ds=‘ls *mat ∣ head -n1’
  set dsbasename=‘basename $ds.nii.gz’
  3dresample \
  -dxyz 2 2 2 \
  -prefix ${dsbasename}.${simulation}.nii \
  -input ${dsbasename}.nii.gz
  End

• 20Sample the E-field NIfTI volumes, using the target ROI to obtain the amount of current likely to be induced, given the corresponding stimulation site and coil handle orientation (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/07.3dmaskave.csh):

  foreach simulation ($sims)
  set ds=‘ls *mat ∣ head -n1’
  set dsbasename=‘basename $ds.nii.gz’
  3dmaskave -q \
  -mask ${subdir}/${subject}/${subject}.rdlpfc.fmask.nii \
  ${dsbasename}.nii > ${statsfile}
  End

• 21
  Draw a set of spheres along each yaw vector and save this as a NIfTI file (

  ${subject}.${simulation}.line.nii.gz

  ) that can be loaded into the neuronavigation software (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/08.3dundump.csh):

  3dUndump \
  -master ${subdir}/${subject}/${subject}_T1fs_conform.nii.gz \
  -prefix ${simulation}.line.nii \
  -datum float \
  -srad 6 \
  -orient RAI \
  -xyz \
  coords.line.1d

• 22Alternatively, write the value corresponding to the amount of current induced in the target region for each yaw vector into a sphere along that yaw vector (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/08.3dundump.csh):

  3dUndump \
  -master ${subdir}/${subject}/${subject}_T1fs_conform.nii.gz \
  -prefix ${simulation}.sphere.nii \
  -datum float \
  -srad 3 \
  -orient RAI \
  -xyz \
  rm.sphere.1d

• 23
  Then combine these files across yaw vectors into a single NIfTI file (

  ${subject}.${simulation}.orientation.mask.nii.gz

  ) using the 3dTstat -max function, which will result in doughnut-shaped mask, tangent to the scalp, at the single-subject target, where the values in the mask reflect the amount of current induced in the target region at each yaw vector (https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/09.save_orientation_masks.csh):

  setenv simulations ‘ls ${hotspot}v*/*sphere.nii’
  3dbucket \
  -prefix ${hotspot}orientation.buck.nii \
  ${simulations}
  3dTstat \
  -max \
  -prefix ${hotspot}orientation.mask.nii \
  ${hotspot}orientation.buck.nii

• 24Export the values corresponding to the current induced in the target region for each yaw vector into a single .csv file (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/10.save_csv.csh).
• 25Create a polar plot by graphing the current induction for each yaw vector as a function of the vector’s direction (Fig. 1f).
• 26Identify the yaw vector that induces the largest E-field in the target ROI and copy the appropriate NIfTI files into an output folder that can be copied to the neuronavigation computer (see https://github.com/balders2/tms_targeting/blob/master/SCRIPTS/11.create_output_directory.csh).

Neuronavigation ● Timing ~15 min

• 27
  Load the high-resolution T1 image (SimNIBS output;

  filename = ${subject}_T1fs_conform.nii.gz

  ) into the neuronavigation software.
• 28
  Load the individual fMRI peak sphere (

  ${subject}.srt-mnt.rdlpfc.nii.gz

  ) into the neuronavigation software.
• 29
  Load the yaw vector line (

  ${subject}.${simulation}.line.nii.gz

  ) and orientation mask (

  ${subject}.${simulation}.orientation.mask.nii.gz

  ) NIfTI files into the neuronavigation software.
• 30Create surfaces for the skin and curvilinear cortical surface using the neuronavigation software.
• 31Create surfaces from the yaw vector line and orientation mask files.
• 32Click on the cortical target to define the x, y, and z coordinates.
• 33Define the roll and pitch vectors as perpendicular to the skin surface.
• 34Define the yaw vector by rotating the orientation of the simulated coil so that it aligns with either the yaw vector line surface or the largest value in the orientation mask surface.
• 35Deliver stimulation to the target region using the vector information provided in Steps 30–34 while the subject performs the designated task (Fig. 2c).

  ? TROUBLESHOOTING

Troubleshooting

See the entries below for troubleshooting advice.

Step 8

It is very important that good alignment exist between the target mask and the T1 images to ensure that the identified BOLD peak lies within the theoretical target region. Accordingly, it is necessary to check mask–T1 alignment before proceeding. A common cause of poor mask–T1 alignment is poor skull stripping. Adjusting the watershed parameters in FreeSurfer can improve skull stripping (see https://surfer.nmr.mgh.harvard.edu/fswiki/FsTutorial/SkullStripFix_tktools).

Step 9

Again, it is critical that there be good alignment between the EPI and T1 images. We recommend using some form of EPI distortion correction (e.g., collecting field maps or reverse phase–encoding blips) and non-linear EPI-T1 registration. It is also important to check EPI–T1 alignment before proceeding. A common cause of poor T1–EPI alignment is a large offset in the initial alignment of these datasets. The afni_proc flag

‘-giant_move’

can often correct this defect.

Step 18

The surface and volume coordinates may be different (e.g., RAS versus surface-RAS). It is important to visually inspect the target once it is loaded into SimNIBS and apply a transformation to the fMRI peak coordinates if necessary.

Step 35

Once the targets have been defined in the neuronavigation software, one last visual inspection should be made to ensure that the correct target region has been identified.

Timing

Steps 1 & 2, target definition: The steps in this section can be done in as little as an hour but will likely require discussion among the study team.

Steps 3–5, data collection: Allow 2 h for consent, instruction, and MRI data collection.

Steps 6–8, structural data processing: Allow at least 24 h for the analytical steps in this section to complete.

Steps 9–11, functional data processing: Allow 2–4 h for the analytical steps in this section to complete.

Steps 12–18, E-field modeling: Allow ~20–30 min per model for the analytical steps in this section to complete. If possible, compute E-field models in parallel to reduce overall time required.

Steps 19–26, target selection: Allow 30 min to complete the steps in this section.

Steps 27–35, neuronavigation: Allow 15 min to complete the steps in this section.

Anticipated results

To demonstrate the scope of the method, we evaluated the scalp-to-cortex distance and intersubject distance of the target, the variability in optimal coil orientation across subjects, and the overall magnitude of the E-field model, given the optimal coil orientation.

In the first analysis, we evaluated the intersubject distance and scalp-to-cortex distance of the single-subject targets on the basis of EPI data obtained during the Sternberg WM paradigm. We contrasted EPI data from the maintenance period of the trials, during which subjects mentally sorted versus maintained a series of five letters (i.e., maintenance versus manipulation).

To constrain our search for these fMRI activation peaks, we used four structural masks obtained using different analytical approaches to identify the bounds of the right dlPFC^49,50. The anatomical mask was obtained for each subject for the middle frontal gyrus ROI, which corresponds to the dlPFC, using the Desikan-Killiany Atlas 41 as output by FreeSurfer (Fig. 3a). The functional mask was obtained using EPI data from the same Sternberg task in an independent sample from a separate study (N = 32; data not shown). For this we extracted the right dlPFC cluster that showed a substantial difference between WM maintenance and WM manipulation (Fig. 3b). The meta-analytical mask was obtained by conducting a keyword search of the NeuroSynth database⁴⁸. For this we used the keyword ‘dlPFC’, conducted a reverse-inference search with a z-score threshold of 7.28, and extracted the cluster centered on the right dlPFC (Fig. 3c). Both the functional and met-analytic mask were normalized to MNI space. The final, unconstrained mask was simply an anatomical mask of the entire right hemisphere (not recommended), included to show the importance of the anatomical masking approach when dealing with single-subject fMRI (Fig. 3d).

Fig. 3 ∣. Example masks obtained using anatomical, functional, meta-analytical, and unconstrained methodologies plotted for a single subject’s T1 scan.

a, Single-subject gray matter mask of the right middle frontal gyrus from the Desikan-Killiany atlas. b, Group mask obtained from subjects performing the Sternberg WM task. c, Mask obtained from the NeuroSynth database from a search on the term ‘Working Memory’. d, Mask of the right hemisphere intended to represent a ‘no mask’ condition. All participants gave written informed consent and were compensated for their time; the study was approved by the National Institute of Mental Health (NIMH) Combined Neuroscience Institutional Review Board.

For the intersubject distance, we first projected all subjects’ T1 images to MNI space, using nonlinear transformations. We then applied these transformations to the coordinates of the target regions so that the targets could be compared across subjects. We then calculated the subject × subject Euclidean distance for targets chosen using each mask (see Fig. 4 for targets and intersubject distance matrices; see Fig. 5 for averages). We then computed a one-way repeated-measures ANOVA comparing these intersubject distances and found a significant main effect (F(3,270) = 138.48; P < 0.001). We then followed this up with pairwise Bonferroni-corrected paired-sample t-tests. First, we found that using any mask results in smaller intersubject distances than using no mask at all (anatomical versus unconstrained: t(90) = 12.193; P < 0.001; functional versus unconstrained: t(90) = 13.827; P < 0.001; meta-analytical versus unconstrained: t(90) = 12.155; P < 0.001). Second, we found that the functional mask resulted in further reductions in intersubject distances compared with the anatomical (functional versus anatomical: t(90) = 2.829; P = 0.035) and meta-analytical (functional versus meta-analytical: t(90) = 3.225; P = 0.011) masks.

Fig. 4 ∣. Peak activations and intersubject distance matrices from Sternberg WM BOLD maps using different masks as input.

a–d, Peak activations from Sternberg WM BOLD maps using anatomical (a), functional (b), meta-analytical (c), and unconstrained (d) masks as input. Peak activations for each subject are pseudo-colored (individual peaks = yellow; overlap = red). e–h, Pairwise intersubject distance matrices (i.e., Euclidean distance between the activation peaks for each pair of subjects) based on coordinates derived using anatomical (e), functional (f), meta-analytical (g), and unconstrained (h) masks as input. All participants gave written informed consent and were compensated for their time; the study was approved by the National Institute of Mental Health (NIMH) Combined Neuroscience Institutional Review Board.

Fig. 5 ∣. Distance metrics of the peak activations from Sternberg WM BOLD maps using different masks as input.

a, Scalp-to-cortex distance of the peak BOLD activation coordinates plotted in Fig. 4. b) Intersubject distance of the peak BOLD activation coordinates plotted in Fig. 4. All participants (N = 14) gave written informed consent and were compensated for their time; the study was approved by the National Institute of Mental Health (NIMH) Combined Neuroscience Institutional Review Board. Bars represent means; error bars represent standard error of the mean.

For the scalp-to-cortex distance, we calculated the Euclidean distance between the cortical target and the projected target for each subject and each ROI (Fig. 5). Given that the E-field attenuates with distance from scalp⁴⁶, this number represents the potential efficacy of the stimulation, given a target. We compared these values across ROIs, using a one-way repeated-measures ANOVA and found a significant main effect (F(3,39) = 11.58; P < 0.001). We then conducted pairwise Bonferroni-corrected paired-sample t-tests to determine which targets were furthest from the scalp. We found that across subjects, all ROIs performed similarly to each other (all P > 0.05) but significantly better than the unconstrained condition (anatomical versus unconstrained: t(13) = 4.058; P = 0.008; functional versus unconstrained: t(13) = 4.321; P = 0.005; meta-analytical versus unconstrained: t(13) = 4.845; P = 0.002), meaning that they had significantly smaller scalp-to-cortex distances.

In the second analysis, we evaluated the ability of this approach to effectively and reliably determine the optimal coil orientation, given a target determined from single-subject BOLD images. We used the targets identified above, from the four mask conditions (anatomical, functional, meta-analytical, and unconstrained), and conducted iterative E-field models with 24 coil handle vectors spread uniformly across the unit circle. We then evaluated these E-field models by sampling the normalized E-field ΔE-field (E/E_max) using either the entire ROI or a 5-mm spherical region centered on the target coordinates.

First, we plotted the normalized E-field as a function of coil handle orientation for all subjects and each ROI–sample combination (Fig. 6). For this analysis, we used both a local sample (i.e., the spherical ROI corresponding to the cortical target) and an ROI sample (i.e., the entire ROI used to define the search region for the peak activation). Either approach could be considered valid, depending on the users’ a priori hypothesis regarding the spatial distribution of the neural activity that they intend to target. For example, if users hypothesize that the target neural process they intend to target is broadly distributed, they should evaluate the E-field models on the basis of a larger ROI (i.e., they should sample the search region). By contrast, if users hypothesize that the neural activity is narrowly distributed, they should evaluate the E-field models on the basis of the amount of current induced locally (i.e., they should sample the spherical ROI corresponding to the cortical target).

Fig. 6 ∣. Changes in the normalized E-field simulated for each target, plotted as a function of coil handle orientation.

a–d, Polar E-field plots for targets derived from anatomical (a), functional (b), meta-analytical (c), and unconstrained (d) masks sampled using a local 5-mm radius sphere. e–h, Polar E-field plots for targets derived from anatomical (e), functional (f), meta-analytical (g), and unconstrained (h) masks sampled using the corresponding ROI mask. Each line represents data from a single subject. Lines are pseudo-colored to enable visual discrimination. Larger distances from the center indicate larger current estimates in the target region at that orientation. All participants gave written informed consent and were compensated for their time; the study was approved by the National Institute of Mental Health (NIMH) Combined Neuroscience Institutional Review Board.

These single-subject polar plots show that the normalized E-field differs as a function of whether it is sampled locally (with a spherical ROI) or at the ROI level, and it can be used to determine the optimal coil orientation for stimulation. Indeed, these polar plots can be viewed on the subject’s T1 scan and loaded into the neuronavigation software (see Step 29). Furthermore, these plots can be used to demonstrate the consistency of the E-field models across subjects. To accomplish this, we identified the optimal coil handle vector for each subject, given a mask condition (anatomical, functional, meta-analytical, or unconstrained) and a sample region (local versus ROI). The goal of this analysis was to determine whether the angle of these vectors was evenly distributed. Because vectors with opposite angles (e.g., 15 degrees and 195 degrees) yielded similar normalized E-field estimates, we averaged estimates for opposing angles to obtain a monomial distribution of optimal vector angles ranging from 0 degrees to 180 degrees. We then tested whether these angles were uniformly distributed along the semicircle for each mask, using a set of Bonferonni-corrected chi-square tests (Fig. 7a).

Fig. 7 ∣. Metrics used to assess E-field models based on input masks and evaluation strategy.

a, Chi-square values representing consistency of coil handle orientation across subjects given distinct input masks (anatomical, functional, meta-analytical, unconstrained) and E-field sampling metrics (local versus ROI). Bars represent chi-square values showing consistency across subjects. b, E-field magnitude given peaks obtained from distinct input masks (anatomical, functional, meta-analytical, unconstrained) based on sample type (local versus ROI). Bars represent means; error bars represent standard error of the mean. All participants (N = 14) gave written informed consent and were compensated for their time; the study was approved by the National Institute of Mental Health (NIMH) Combined Neuroscience Institutional Review Board.

When we sampled the normalized E-fields, using the mask regions, we found that targets selected from within the functional mask and the meta-analytical mask showed a significantly non-random distribution of optimal coil handle angles (functional: χ²(13) = 28.86; P = 0.027; meta-analytical: χ²(13) = 27.14; P = 0.048). However, targets selected from within the anatomical mask and the unconstrained mask conditions seemed to show a random distribution of optimal coil handle angles (P > 0.05). When we sampled the normalized E-fields using the local spherical ROI, we found that no mask condition yielded a significantly non-random distribution of optimal coil handle angles (all P > 0.05). These results suggest that the optimal coil handle orientation is consistent across subjects when either the functional mask or the meta-analytical mask is used to select the location of the TMS target and that this consistency is observed only when the normalized E-field model is sampled with the ROI mask.

Finally, we estimated the overall magnitude of the E-field model, given the optimal coil handle vector angle, and plotted this as a function of mask and sample condition (Fig. 7b). We analyzed these data using a 2 (local versus ROI) × 4 (anatomical, functional, meta-analytical, versus unconstrained) repeated-measures ANOVA. We found a significant main effect of sample, suggesting that the normalized E-field is larger when sampled locally (F(1,13) = 76.71; P < 0.001). We also found a significant interaction (F(3,39) = 9.66; P < 0.001), which we probed using pairwise Bonferroni-corrected paired-sample t-tests. For samples taken using the local spherical region, there were no differences among the mask conditions (all P > 0.05). However, for samples taken using the entire ROI, the normalized E-field was larger when sampled with all three of the other masks, compared with the unconstrained mask condition (i.e., the right hemisphere; anatomical versus unconstrained: t(13) = 11.967; P < 0.001; functional versus unconstrained: t(13) = 18.172; P < 0.001; meta-analytical versus unconstrained: t(13) = 11.164; P < 0.001).

Data availability

Example data have been uploaded as Supplementary Data.