This repository includes algorithms in MATLAB for localizing intracranial contact. CT images, MRI images, and freesurfer software are required.
Updates: 2025 Aug 19, I developed an automated pipeline for estimating intracranial electrode location based on CT images and manufacture information of the electrode lead. The main script has been tested on DIXI electrodes.
git clone https://github.com/GanshengT/intracranial_contact_loc.git
cd intracranial_contact_locElectrode localization refers to identifying the precise anatomical positions of implanted intracranial electrodes.
This step is essential in both clinical practice and neuroscience research:
-
Clinical applications:
- Mapping eloquent cortex for surgical planning.
- Delineating seizure onset zones in patients with epilepsy.
- Guiding resection or stimulation strategies with minimal risk.
-
Research applications:
- Linking electrophysiological signals to anatomical structures.
- Understanding network-level dynamics of cognition and disease.
- Enabling reproducible and anatomically grounded scientific discovery.
Precise and objective localization is thus critical. However, most workflows still rely heavily on manual localization, which presents two major limitations:
- Subjectivity – human raters must distinguish contacts on CT scans that often suffer from streak artifacts and noisy intensity. For example, an isosurface in FreeSurfer may appear as a continuous lead, making individual contacts ambiguous.
- Time-consuming and error-prone – manual tracing and labeling are labor-intensive, and mistakes can propagate through analysis pipelines.
To address these issues, v1.0 of this pipeline introduced an automated approach that leverages CT intensity and a connected-graph searching algorithm to identify electrode contacts.
Building on this foundation, v2.0 incorporates electrode manufacturing specifications to further improve robustness. By exploiting the known geometry of linear electrode leads, the algorithm can fit multiple contacts simultaneously, mitigating the impact of noisy voxels and imaging artifacts. This allows faster, more accurate, and reproducible electrode localization for both clinical and research use cases.
Our automated electrode localization approach begins by analyzing the CT intensity distribution.
We identify peaks corresponding to metal contacts, which allows us to select candidate voxels associated with the electrode lead.
Figure: Example CT intensity distribution. We find peaks with prominance higher than the noise. The right peak correspond to metal voxels
We visualize the metal voxel isosurface to sanity-check the threshold and the spatial distribution of metal (contacts + lead body). We also overlay the planned trajectory (exported from ROSA) and restrict analysis to voxels that lie within a cylindrical neighborhood of this line segment. This (i) suppresses unrelated metal (e.g., screws/plates) and (ii) disambiguates adjacent leads.
Note on ROSA coordinates: Ensure the ROSA trajectory is in the same RAS frame as the CT used for localization. Co registration is needed
Click the image to watch the rotation video.
Figure: Extracted metal voxels and the planned trajectory for one lead
From the set of metal voxels near the planned trajectory, we employed a seed-growing algorithm to isolate the voxels corresponding to the intended lead. The main objective is to disambiguate this lead from neighboring electrodes or unrelated metal.
To achieve this, multiple seeds are initialized within the metal cluster, and each seed iteratively expands its voxel pool. At each step, a neighboring voxel is added if it lies close to the centroid of the current pool, thereby maintaining a coherent and compact structure.
To prevent growth onto adjacent leads, we evaluate the filling ratio of a cylindrical tube aligned with the planned trajectory and encompassing the current voxel pool. Because different leads typically have offsets or angular deviations relative to the planned path, including voxels from another lead would substantially reduce the cylinder’s filling ratio. This provides a natural safeguard against erroneous extension across multiple leads.
Furthermore, we identify the proximal entry point and the most distal point based on segmenting the brain from the CT. We checked the interception between the identified lead voxel and the brain mask and check which direction (between the two end points, that is, the interception point and the distal point) align with the boundary to center vectors from the brain mask.
In rare occasion, two trajectories were very close. For clinical reason, it is usually not ideal to have two parallel leads that are close (because the regions are already covered by one lead).
Therefore, we implement a novel algorithm based on global coherence. Global coherence is defined as the cosine similarity of directions for vectors between a voxel of interest and other voxels (that are further from this voxel, which is identified by distance percentile). In this scenario, voxels that belong to one lead will have higher global coherence as the vectors between this voxel and other voxels follow the lead direction.
After modeling the lead, we calculate electrode likelihood along the modelled line/curve. An electrode can be modelled by a cylinder with radius and length. For each location, we subtract the background (outter shell) CT intensity from the elecrtode cylinder intensity as the electrode likelihood.
Most of the time, the lead is presented as segmented line in the CT images, where electrodes are separated. Below is an example of a typical lead in CT images.
We have updated the method for estimating the most probable location for electrodes. The main update is defining a core and shell, the most probable location for an electrode is the location where the voxel intensity in the core is greater than the voxel intensity in the shell. The core is modeled to be a cylinder having the dimension provided by the manufacture The shell is defined to be a cylinder encompassing the core, which cover the insulating space (space between two electrodes), and the annulus surrounding the core (to capture the streak artifact resulting from metals in CT imaging)
This method improves the localization of electrode contacts even when CT images are noisy (aka, the scattering of the electrode obscures the boundary between electrodes, making the lead a non-segmented streak).
We also created new function for localization of DIXI MICRODEEP® MICRO-MACRO DEPTH ELECTRODES This is done by locating all macro-contacts and retain the functioning ones based on the electrode types: 6 macro-contacts / 3 tetrodes (12 micro-contacts) 9 macro-contacts / 2 tetrodes (8 micro-contacts).
One can plug in this function in existing Imaging processing pipeline by calling gt_automatic_localize_electrode.m. A minimal working example called minimal_working_example.m shows how to properly provide input to the function. An example of integration is the successful integration into VERA.
Please note that CT scans in NIFTY format and electrode specification are required.
Additionally, for windows users, zipped imaging files xxxxx.mgz are not supported by MRIread function from freesurfer (which is used in this external wrapper function). Therefore, please provide unzipped imaging files xxxxx.mgh as input.
Please use this link (https://ganshengt.github.io/intracranial_contact_loc/) for detailed documentation.