iEEG-recon: A Fast and Scalable Pipeline for Accurate Reconstruction of Intracranial Electrodes and Implantable Devices

Abstract

Objective:

Clinicians use intracranial EEG (iEEG) in conjunction with non-invasive brain imaging to identify epileptic networks and target therapy for drug-resistant epilepsy cases. Our goal was to promote ongoing and future collaboration by automating the process of “electrode reconstruction,” which involves the labeling, registration, and assignment of iEEG electrode coordinates on neuroimaging. We developed a standalone, modular pipeline that performs electrode reconstruction. We demonstrate our tool’s compatibility with clinical and research workflows and its scalability on cloud platforms.

Methods:

We created iEEG-recon, a scalable electrode reconstruction pipeline for semi-automatic iEEG annotation, rapid image registration, and electrode assignment on brain MRIs. Its modular architecture includes: a clinical module for electrode labeling and localization, and a research module for automated data processing and electrode contact assignment. To ensure accessibility for users with limited programming and imaging expertise, we packaged iEEG-recon in a containerized format that allows integration into clinical workflows. We propose a cloud-based implementation of iEEG-recon and test our pipeline on data from 132 patients at two epilepsy centers using retrospective and prospective cohorts.

Results:

We used iEEG-recon to accurately reconstruct electrodes in both electrocorticography (ECoG) and stereoelectroencephalography (SEEG) cases with a 30-minute running time per case (including semi-automatic electrode labeling and reconstruction). iEEG-recon generates quality assurance reports and visualizations to support epilepsy surgery discussions. Reconstruction outputs from the clinical module were radiologically validated through pre- and post-implant T1-MRI visual inspections. We also found that our use of ANTsPyNet deep learning-based brain segmentation for electrode classification was consistent with the widely used Freesurfer segmentations.

Significance:

iEEG-recon is a robust pipeline for automating reconstruction of iEEG electrodes and implantable devices on brain MRI, promoting fast data analysis, and integration into clinical workflows. iEEG-recon’s accuracy, speed, and compatibility with cloud platforms makes it a useful resource for epilepsy centers worldwide.

Graphical Abstract

Introduction

Treatment of drug resistant epilepsy often requires precise localization of epileptic networks responsible for seizures. To achieve this, clinicians commonly use a combination of non-invasive and invasive techniques. Among them, intracranial electroencephalography (iEEG) stands out due to its ability to provide direct measurements of brain activity with high temporal resolution. After implantation of electrodes in the brain, however, it is necessary to confirm their anatomical location. This ensures effective sampling during the study, interpretation of brain patterns, and planning for further treatment. To determine electrode positioning post-surgery (i.e., “electrode reconstruction”) the process involves electrode labeling, CT and MRI co-registration, and the systematic assignment of electrode contact to specific brain regions. Despite its clinical significance, there remains an unmet need for fast, accurate, and easily implementable iEEG electrode reconstruction tools for clinical workflows.

Various techniques have been proposed for iEEG electrode reconstruction^2–7. While these methods have been useful in research for determining brain areas that produce epileptiform activities and assessing overlap with the location of surgery, their clinical adoption remains limited. The slow adoption is primarily due to three challenges. First, available tools have a steep learning curve for users unfamiliar with programming and image processing. Second, current tools are often time-consuming (~4+ hours runtime), and require constant user input, which makes them impractical for high-volume centers or large multicenter prospective clinical trials. Scalability necessitates consistent and validated pipelines that can process hundreds of cases on both centralized and federated platforms, while ensuring appropriate data harmonization. Finally, the currently available pipelines do not leverage recent deep learning-based improvements in image registration and segmentation techniques, which reduce long runtimes and computational requirements, while increasing accessibility.

In this study, we developed and validated iEEG-recon, a standalone pipeline for iEEG electrode reconstruction. We achieved this by a) semi-automatically marking the electrodes on post-implant computer tomography (CT) images, b) co-registering post-implant CT to pre-implant MRI using state-of-the-art rapid brain segmentation and co-registration techniques, and c) incorporating a modular, scalable design consisting of core modules for clinical needs and research modules for flexible parameter tuning. We developed versions of our pipeline compatible with MATLAB and Python programming environments, as well as a stand-alone containerized tool that can be deployed on cloud-based infrastructure. We validated our pipeline retrospectively on extensive patient datasets containing both stereotactic EEG (SEEG) and electrocorticography (ECoG) from post-implant MRI, then tested its viability on data collected prospectively from two level-4 U.S. epilepsy centers. For each run, iEEG-recon generates a quality assurance report and visualizations that can facilitate discussion in epilepsy surgery meetings. Our pipeline is fast, scalable to process hundreds of patient datasets, reproducible, and available as an open-access tool for wider adoption.

Methods

Participants

We included 132 patients with drug-resistant epilepsy from two epilepsy centers: the Hospital University of Pennsylvania (HUP: n = 109) and the Medical University of South Carolina (MUSC: n = 23). Our methods were developed and validated on a retrospective cohort of 118 patients (HUP: n = 98 and MUSC: n = 20) and implemented for prospective testing on 14 patients (HUP: n = 11 and MUSC: n = 3). Patients were enrolled serially between 2015 and 2023 after providing written informed consent for iEEG data analysis, in line with the University of Pennsylvania’s IRB-approved protocol (reference number 821778). All patients underwent whole-brain MRI and iEEG implantation (ECoG: n = 23 or SEEG: n = 75), followed by post-implant CT scans. The retrospective cohort had detailed clinical annotations from presurgical evaluation meetings and follow-up records on surgical outcomes. Multiple surgical procedures were represented in our dataset, including ablation (N = 33), surgical resection (N=34), responsive neurostimulation (N=9), deep brain stimulation (N=1), vagus nerve stimulation (N=1), and no surgery to date (N=18). The prospective cohort data were processed before epilepsy surgery meetings by clinical coordinators and fellows to assess ease of use in routine clinical workflows. Feedback from an expert neuroradiologist (J.S.) was obtained to improve clinical reporting; however, iEEG_recon outputs were not used in clinical decision-making. Subject demographics and clinical characteristics are summarized in Table 1.

Table 1 -.

Subject demographics and clinical characteristics

Subject Demographics
Characteristic	Penn - Retrospective	Penn - Prospective	MUSC
Total Subjects	98	13	23
Age	34±11	37±12	35±15
Female	42	9	10
Disease Duration (years)	15±13	16±13	15±13
Lesional MRI	40	8	10
Disease Laterality
Left	47	4	6
Right	32	6	9
Bilateral	11	2	4
Unknown	8	1	4
Type of Implant
ECoG	23	0	0
SEEG	75	13	23
Type of Surgery
Ablation	33	3	1
Resection	34	1	5
RNS	9	1	3
DBS	2	1	0
VNS	2	0	0
No Surgery	18	7	14

Image Acquisition

At both centers (HUP and MUSC), MRI data were collected on a 3T Siemens Magnetom Trio scanner using a 32-channel phased-array head coil prior to electrode implantation. Anatomical images were acquired using a magnetization prepared rapid gradient echo (MPRAGE) T1-weighted sequence (HUP: repetition time = 1810ms, echo time = 3.51ms, field of view = 240mm, resolution = 0.94×0.94×1.0 mm3; MUSC: repetition time = 1900ms, echo time = 2.36ms, field of view = 256mm, resolution = 1.0×1.0×1.0 mm3). Following electrode implantation, spiral CT images (Siemens) were obtained clinically for the purposes of electrode localization. Both bone and tissue windows were obtained (120kV, 300mA, axial slice thickness = 1.0mm, same for both institutions). We converted the dicom files from scanner to NIFTI format using the dcm2niix converter (https://www.nitrc.org/projects/mricrogl).

Image Processing

An overview of the implant reconstruction pipeline is shown in Figure 1. We separated our tool into three sequential modules to allow users to choose the level of processing for their use case.

Figure 1 – Pipeline Overview:

The iEEG-recon pipeline offers a comprehensive solution for electrode marking and reconstruction. By leveraging semi-automatic electrode identification on post-implant computer tomography (CT) images (Module 1), it simplifies and accelerates the process. The pipeline further enhances accuracy by co-registering the post-implant CT with pre-implant MRI using state-of-the-art rapid brain segmentation and co-registration techniques. Built with versatility in mind, iEEG-recon encompasses both core modules tailored for clinical needs (Module 2) and research modules (Module 3) for flexible parameter tuning. This stand-alone and containerized tool can be effortlessly deployed on cloud-based infrastructure.

Module 1 – VoxTool for Electrode Labeling

In our pipeline, electrode labeling is conducted using VoxTool, a user-friendly graphical user interface for electrode labeling. This software was developed in collaboration with the Penn Memory Lab (https://memory.psych.upenn.edu/Main_Page). Briefly, post-implant CT scans are loaded into VoxTool, which applies an intensity threshold to accentuate the electrodes. The user manually enters electrode labels, then navigates a 3D viewer where electrode voxels can be selected by directly clicking on them. Users can then interpolate electrode labels for grid, strip, and depth electrodes with a single click. We have created a video tutorial that demonstrates how to use VoxTool for loading the 3D graphical user interface (GUI) and semi-automatically annotating electrodes. To access the tutorial, please follow this link: https://github.com/penn-cnt/ieeg-recon/tree/main/voxTool.

Module 2 – Post-implant CT to Pre-Implant MRI Registration

The objective of Module 2 is to register the post-implant CT to the pre-implant MRI. To ensure appropriate skull matching between the CT and MRI acquisitions, we first threshold the intensity of the post-implant CT scan such that only the skull and electrode contacts are visible. After CT thresholding, a rigid registration procedure with 6 degrees of freedom (3 rotation and 3 translation) is applied between the thresholded post-implant CT and the pre-implant MRI. Since brain and skull shapes should not change drastically pre- and post-implantation, 6 degrees of freedom are sufficient for accurate registration^3,4,6. Our pipeline has the option to use either FLIRT^8,9 or Greedy¹⁰ (https://github.com/pyushkevich/greedy) for registration. FLIRT parameters are: 640 histogram bins to discretize the intensity values of the images being registered, and a mutual information cost function with 6 degrees of freedom for registration. For Greedy, we initialize the registration at the center of mass of each image and apply a multi-resolution registration process with 100 iterations at the highest resolution and 100 iterations at the second highest resolution. We stopped the registration after the second highest resolution stage as we did not see significant improvements in registration accuracy by including additional lower resolution levels, but runtime increased drastically. Greedy registration utilizes normalized mutual information as the cost-function with 6 degrees of freedom. We recommend using Greedy as it is faster, but in cases where it fails, FLIRT is available as a fallback option.

The post-implant CT to pre-implant MRI registration generates a transformation matrix that transforms the original electrode coordinates from the post-implant CT space to the pre-implant MRI space. The results of the registration are compiled in an HTML report for quality assurance. For visualization of reconstructed electrodes, an interactive workspace file is automatically generated for visualization in ITK-Snap (www.itksnap.org)¹¹ (Figure 2), an open-source application to visualize and manipulate biomedical images. These visualizations are necessary for manual quality assurance, and any residual rotation between scans could be easily corrected within this interface.

Figure 2 – Quality Control Files for Module 2:

The output of Module 2 generates an HTML report (A.) that shows the accuracy of the registration between the pre-implant MRI and the post-implant CT, as well as a 3-dimensional scatterplot of the electrodes from the thresholded CT scan and their manually labeled coordinates in the pre-implant MRI space. ieeg_recon also generates an ITK-Snap workspace file that overlays the thresholded post-implant CT and pre-implant MRI in the pre-implant MRI space, and plots the coordinates of each of the electrodes as well as their electrode labels provided by VoxTool. This workspace allows for interactive visualization and quality assurance of the electrodes and their locations.

While this process is described for registering the post-implant CT to the pre-implant MRI as a reference, the same process can be applied to any other pair of images (MRI or CT) if one is chosen as a reference. For example, our pipeline also works if the post-implant MRI is registered to the pre-implant MRI, or the post-surgical MRI is registered to the pre-implant MRI (see below).

Module 3 – Electrode Region-of-Interest Assignment

Pre-implant MRI Segmentation

The brain can be subdivided into contiguous regions of interest (ROIs) based on the structural or functional similarity within each region. Assigning electrodes to specific ROIs is useful for confirming iEEG targets¹². To assign electrodes to a specific ROI, we must identify ROIs on the reference pre-implant MRI. Our pipeline can take as an input any atlas–a particular parcellation scheme–that is in the same space as the reference pre-implant MRI, including: (1) cortical parcellations and subcortical segmentations from either the Desikan-Killany-Tourville (DKT) atlas^13,14 or the Taliarach atlas¹⁵ from FreeSurfer; (2) standard atlases in MNI space (AAL¹⁶, Schaffer¹⁷, etc) pre-registered to the reference pre-implant MRI; (3) outputs from more specialized subcortical segmentation approaches such as ASHS¹⁸ and THOMAS¹⁹. Our pipeline enables users to choose any atlas for electrode assignments and encourages application of multiple atlases for robustness²⁰.

ANTsPyNet Segmentation

IEEG-recon has an option to generate and use a cortical-subcortical DKT parcellation, and tissue segmentation (DeepAtropos), from pre-implant MRI using ANTsPyNet²¹. We included ANTsPyNet, as opposed to FreeSurfer, due to its speed, generating both cortical and subcortical segmentations in ~5 minutes. If the user has a FreeSurfer segmentation they would like to use, as specified in the previous section, the pipeline allows it as an input.

Electrode Region of Interest Assignment

Once an atlas has been chosen, the electrodes transformed into pre-implant MRI space can then be assigned to the specific ROIs in the pre-implant MRI. To do so, we generate a sphere with a user-defined radius around each electrode contact coordinate obtained from Module 2. The percent overlap of this sphere with the atlas brain regions is computed, and the region with the most overlap is the one assigned to that electrode. This process is repeated for each electrode. Besides identifying the region with the highest overlap, the software also generates a ranked list of percent overlaps when an electrode comes into contact with multiple regions. This feature enables users to choose a specific sphere-ROI overlap threshold that best aligns with their research question, offering more flexibility in the analysis process.

Template Space Registration

We provide the option for obtaining coordinates in MNI space after an affine registration applied between the native space reference MRI and the MNI152NLin2009cAsym template. This transformation applies the same robust non-linear registration based on antsRegistration that is built into the anatomical preprocessing pipeline of fMRIprep⁴³.

We also included 16 standard space atlases to provide users with the option to reconstruct electrodes to a standard atlas of their choice. Contacts localized to a region in native space should correspond to the same region in normalized space. However, the application of non-linear registration warps can shift the electrode position, potentially localizing it to a different region compared to where the contacts were localized in native space. To circumvent this issue, our implementation applies the registration to transform the standard space atlases to native space, ensuring accurate localization regardless of the atlas chosen.

Brain-shift correction

iEEG-recon can be used for both ECOG and SEEG implantations. Notably, with ECOG implants, brain shifts are common due to the inward displacement of brain tissue and electrodes. This displacement results from pressure changes associated with the craniotomy. Consequently, the original electrode reconstruction can appear submerged within the parenchyma, rather than on the cortical surface. To address these shifts, we have implemented a two-step non-linear optimization technique which has been previously demonstrated to correct for brain shifts. This approach uses an energy-minimization algorithm to preserve the inter-electrode distance while repositioning the displaced electrodes onto the pial surface⁴⁰. The supplementary methods provide a detailed description of this algorithm.

To run brain-shift correction in iEEG-recon cortical surfaces are required. For both the command line interface and GUI, a user can point the location of the Freesurfer directory (recon-all output) to automatically run brain-shift correction. The pipeline only applies the brain shift correction to those electrodes labeled as grids or strips within VoxTool. Depth electrodes are left unchanged.

Post-Surgery and Post-Implant Registration

Our pipeline additionally provides the option for registering a post-surgery MRI (e.g. an MRI with a surgical resection cavity) to the original pre-implant MRI. This is useful to detect the electrodes that were implanted in resected brain areas. The rigid registration between the post-resection MRI and the pre-implant MRI uses the same Greedy parameters as the one used for registering the post-implant CT to pre-implant MRI. An example output from this post-resection registration pipeline is shown in Figure 3. Following similar steps, our pipeline can also register the post-implant MRI to the pre-implant MRI.

Figure 3 – Estimating the original electrodes located in the post-surgical resection mask:

Pre-resection MRI for an example subject who underwent resective surgery (A.), post-resection MRI (B.), and the Module 3 pipeline output identifying electrode contacts within the resected brain region (C.). The green region represents the resection cavity mask, the blue dots represent the electrodes within the resection cavity, the red dots represent the electrodes outside the resection cavity.

Running iEEG-recon

We provide detailed documentation on how to run iEEG-recon in both Matlab and Python (https://ieeg-recon.readthedocs.io). To run the pipeline, the user will need the following minimum requirements: labeled electrode coordinates from VoxTool, a pre-implant MRI, and a post-implant CT. The process also utilizes a BIDS-like structure, which organizes the data for each subject according to subject IDs (e.g., sub-PENN01) and sessions (e.g., ses-research3T, ses-clinical01) using a BIDS-like naming convention. Semi-automatic electrode labeling with VoxTool takes between 20–35 minutes with a 10–16 electrode contact surgical plan. The entire iEEG-recon pipeline can be run with an intuitive Graphical User Interface (GUI), which includes VoxTool, a File Creator (for generating the required folder structure from source data), and one-click options for Module 2 and 3 (Supplementary Figure 3).

Cloud Implementation

To scale clinical and research workflows, we also developed a cloud implementation of iEEG-recon leveraging a containerized version of the pipeline. To automate the iEEG-recon pipeline on a cloud infrastructure, we deploy our code on an AWS EC2 (Amazon Elastic Compute Cloud) instance. For our purposes, we regard an EC2 instance (https://docs.aws.amazon.com/ec2/) as a virtual machine capable of running the dockerized iEEG-recon software consistently across all collaborators. This approach simplifies the process and ensures uniformity, regardless of where the collaborator is based. To help ensure patient anonymity, we have additionally included a defacing option to iEEG-recon, capable of removing facial features from both inputs to the pipeline and outputs from the pipeline. Details of the cloud implementation and the defacing approach can be found in the supplementary materials.

Results

Registration and Electrode ROI Assignment Within Minutes

One of the advantages of our pipeline is that the entirety of Modules 2 and 3, that is, the post-implant CT to pre-implant MRI registration and the electrode ROI assignment, run in about 10±4 minutes if Greedy is used for registration and ANTsPyNet is used for segmentation (default options). We tested the run time on standard laptops with a minimum of 8GB of RAM. This allows for fast turnaround times in situations where electrode reconstruction results are needed urgently (e.g. when a patient is being presented in a surgical conference), or when many subjects are being processed simultaneously from raw input images.

Reconstructed Electrode Locations

Reconstructed electrode locations in MNI space (Figure 4) demonstrate a spatial bias for the temporal lobes, and specifically, the left temporal lobe. This is consistent with our studied population of mostly temporal lobe epilepsy patients. Medio-dorsal and medio-ventral structures were rarely implanted in our cohort.

Figure 4 – Reconstructed electrode locations in MNI template space:

Electrode locations in MNI template space across all 98 subjects. Electrodes are represented as dots overlaid on an MNI surface template.

Across all patients, 9.2 ± 8.4% of the labeled electrodes were found to be in no tissue (e.g. outside the brain), 14.4 ± 8.5% in the CSF, 32.7 ± 11.7% in the gray matter, and 42.6 ± 13.5% in the white matter when using a 2mm radius for region assignment (Supplementary Figure 1A). By default, ieeg_recon assigns regions based on a majority voting procedure, where the region with the most overlap with a 2mm sphere around the electrode radius is assigned to that electrode. However, sometimes brain shifts and post-surgical swelling can slightly alter the assignment of the electrodes. To overcome this, we provide a detailed JSON file that describes the percent overlap of each electrode sphere with all the regions it overlaps with, allowing for re-assignment of electrodes to different regions depending on different thresholds of overlap. In Supplementary Figure 1B–E we show how the distribution of electrodes changes if we re-assign electrodes originally assigned to either white matter, CSF or outside the brain, to a gray matter region if their overlap with that region was at least as large as the specified threshold.

Among the gray matter electrodes, we found that the top 3 implanted regions were the left middle, inferior and superior temporal gyri, with 71/98, 59/98 and 54/98 subjects having those regions implanted (Supplementary Table 1). Similarly, when counting the number of gray matter electrodes implanted in each region, 373/4636 were found in the left middle temporal gyrus, 335/4636 in the left superior temporal gyrus, and 215/4636 in the left inferior temporal gyrus (Supplementary Table 2).

Our pipeline was robust to patients with prior surgery. Four of the patients included in our cohort had a prior resection, and our pipeline was able to register and reconstruct the electrode locations in these patients. We have included examples of these patients in Supplementary Figure 5.

The electrode localizations for all patients were validated by visual inspection of post-implant MRI by a board-certified neuroradiologist (J.S.), which provides confidence in the accuracy and reliability of our pipeline.

ANTsPyNet Electrode ROI Assignments are Comparable to FreeSurfer

The deep learning-based DKT segmentation provided by ANTsPyNet allows ieeg_recon to run quickly, while still providing accurate electrode ROI assignments. However, the most common approach for individualized brain segmentation in electrode reconstruction is to use FreeSurfer derived DKT segmentations^2–7. We tested how consistent the electrode region assignments were between FreeSurfer and ANTsPyNet segmentations. At a group level, we found that the number of patients with electrodes implanted in each region was highly correlated between ANTsPyNet and FreeSurfer segmentations (Pearson’s r=0.96) (Figure 5). A similar correlation is seen if the number of electrodes assigned to each region across all patients is counted. However, the FreeSurfer implementation took an average of 5 hours on a dedicated computing server, whereas the ANTsPyNet approach was completed in an average of 10 minutes on the same server.

Figure 5 – FreeSurfer and ANTsPyNet Desikan-Killany-Tourville atlas electrode assignments across subjects:

Panels A-B. show the number of subjects for which an electrode was assigned to each of the regions in the Desikan-Killany-Tourville (DKT) atlas when the atlas segmentation was done by FreeSurfer (A.) or ANTsPyNet (B.). Panel C shows the correlation of that count across ROIs for the FreeSurfer and ANTsPyNet segmentations. r – Pearson’s r

iEEG-recon is robust to implantable device artifact

Nine of the 98 patients were implanted with a responsive neurostimulation (RNS, NeuroPace Inc.) device after their iEEG implantation. The RNS computer is housed in a titanium case that is implanted within the thickness of the skull, and two electrode leads are implanted in the tissue of the brain. We used ieeg-recon to localize both the RNS electrode contacts and location of the device case by registering a post-RNS-implant CT scan to the pre-implant MRI. The post-RNS-implant CT scan presented a unique registration challenge due to the device artifact in the skull, since the registration steps operate by aligning the contours of the skull between CT and MRI images. For all RNS patients, we found that the pipeline outputs were acceptable as verified by visual inspection using our quality assessment reports and confirmed by a board-certified neuroradiologist (J.S.). A representative example of this reconstruction is shown in Supplementary Figure 2. The successful electrode localization in these RNS patients demonstrates the robustness of our reconstruction pipeline in handling multiple types of potential imaging artifacts.

Discussion

We present iEEG-recon, an innovative electrode reconstruction pipeline that leverages cutting-edge image registration and segmentation algorithms for rapid and precise intracranial electrode localization. iEEG-recon’s scalable modular design enables users to select the most suitable options for their research objectives or clinical applications. We evaluated and tested iEEG-recon in a cohort of retrospective and prospective patients with drug-resistant epilepsy who underwent iEEG implantations at two epilepsy centers. The tool’s accuracy, speed, and compatibility with cloud platforms, demonstrated through its deployment on the Flywheel and Pennseive data infrastructure, make it a valuable resource for epilepsy centers worldwide. Furthermore, iEEG-recon proved robust against artifacts caused by RNS, highlighting its utility for patients with intracranial devices. In summary, iEEG-recon demonstrates the first step towards developing, validating, and testing scalable quantitative tools for standardized data analysis and seamless integration into clinical workflows, for advancing epilepsy treatment through multicenter collaborations.

Current electrode reconstruction pipelines^3,4,7 have several limitations that hinder their widespread adoption across clinical sites that we believe iEEG-recon addresses. First, there is a steep learning curve for many of these tools, requiring the user to perform scripting tasks or to directly write code in a specific programming language in order to run the pipeline. iEEG-recon is self-contained, and as long as the input files are in the correct format and the electrodes have been labeled, the rest of the pipeline runs with a single command. Second, many of these tools are not scalable, as they require complex setups, and sometimes closed source software, limiting flexibility and modularity. Providing a Docker container allows iEEG-recon to be executed in any machine or server capable of running Docker, allowing the exact same software to be executed across multiple clinical or research sites. Finally, many currently available pipelines make use of effective, yet old, registration and segmentation technology, which leads to high computational requirements and running times. By leveraging advances in image registration and deep-learning based segmentation, iEEG-recon allows for complete electrode reconstruction and region-of-interest electrode assignments within 10–15 minutes, a significant improvement relative to current approaches. While commercial solutions exist for electrode localization (e.g. CURRY), these solutions are often costly, making them available mostly to resource-rich epilepsy centers. iEEG-recon provides the key functionalities available in most of these commercial pipelines in an open-source package.

Electrode reconstruction remains a critical component of presurgical epilepsy planning²², and iEEG-recon can help streamline and improve this process. iEEG-recon could be used to accurately localize intracranial electrodes in patients with refractory epilepsy, providing important information for surgical planning, such as where the seizure-onset zone (SOZ) electrodes are located²³, and whether there is overlap with inoperable structures such as the eloquent cortex. After a patient has undergone epilepsy surgery, iEEG-recon could be used to confirm whether the post-surgical resection site overlaps with the suspected SOZ electrodes. iEEG-recon can also be used to identify the location of RNS electrodes for post-surgical confirmation. Similar to iEEG-recon, the recently proposed SEEGAtlas plugin for the IBIS system represents another open-source software platform for image-guided neurosurgery and multimodal image visualization, which could be particularly useful in combination with pre- and post-implant T1 MRI^41,42. Taken together, these streamlined tools emphasize the need to optimize electrode implantation planning to assist neurosurgeons and neurologists in planning IEEG implantation to interpret iEEG data.

Electrode reconstruction is also critical for research that involves intracranial electrophysiology²⁶. Many intracranial electrophysiology studies focus on specific brain regions, such as the hippocampus^27,28 and the motor cortex^29,30, in order to understand function in health and disease. Therefore, accurate identification of electrodes within the target regions is necessary to generate reproducible findings, and to ensure that the measured signal is associated with the structure of interest. Multimodal studies that combine iEEG with neuroimaging, such as diffusion tensor imaging, are also of interest across disciplines, including epilepsy^12,31,32. Functional neuroimaging, such as resting-state fMRI, has also demonstrated widespread abnormalities in focal epilepsy^33–35, yet direct intracranial electrophysiological correlates of these abnormalities are lacking. iEEG-recon provides a natural framework for bridging intracranial electrophysiology and neuroimaging by allowing different structural and functional neuroimaging-based atlases to be used in the electrode reconstruction process.

The use of federated data analysis has emerged as a promising approach in neuroimaging and iEEG studies, enabling researchers to perform large-scale multi-center analysis while maintaining data privacy^36–38. We built iEEG-recon to work with standardized data formats (BIDS) and to streamline the processing of neuroimaging data, reducing the need for specialized technical expertise and enabling researchers to focus on interpreting their results. By improving the accessibility of neuroimage processing, our pipeline can facilitate the collaboration of multiple research teams across centers, ultimately leading to more robust and generalizable findings³⁹.

There are several limitations of our current work. First, while iEEG-recon has been successfully tested in two level 4 adult epilepsy centers, its efficacy in pediatric populations remains unknown. Second, while self-contained, iEEG-recon leverages advances in image registration and segmentation algorithms, which have been sourced from other toolboxes that may still be under development. While these advancements have enabled the tool to provide more efficient results, it is important to consider the potential limitations of these technologies. Finally, it is important to mention that iEEG-recon is currently not FDA approved, and further testing and optimization are needed before it can be implemented reliably in clinical practice. However, the authors invite collaboration and feedback from the scientific and medical communities to improve and refine the tool’s performance.

Conclusion

Reconstructing iEEG electrodes and precisely localizing them is necessary for research and clinical applications in epilepsy. A primary goal of our open-access tool is to make electrode reconstruction accessible to those with limited computing background or resources. Future work should iterate upon our open-source pipeline, and we expect improvement of our tool with widespread use and feedback.

Key Points:

1. iEEG-recon is a modular scalable pipeline to reconstruct electrodes contacts from iEEG and implantable devices.

2. iEEG-recon features a graphical user interface (GUI), command line interface (CLI), cloud implementation, and detailed quality assurance reports.

3. iEEG-recon is validated on both retrospective and prospective multicenter epilepsy data.

4. iEEG-recon is highly practical for implementation in routine clinical workflows and research applications that require accurate electrode reconstruction.

Data Availability Statement

The entire iEEG-recon pipeline and details for usage can be found here: https://ieeg-recon.readthedocs.io/en/latest/