Abstracts

Rationale: More than one million Americans suffer from drug-resistant epilepsy, and, of these, 50-60% would benefit from surgical removal of the epileptogenic focus. Invasive electrophysiology is often utilized for seizure localization and functional mapping prior to surgery to maximize seizure control while minimizing the risk of deficits. Accurate localization of implanted electrodes with respect to patient-specific anatomy is essential to interpret electrocorticographic data; however, current localization methods require manual intervention, which is time-consuming and susceptible to errors. We present a fully automated solution to electrode localization and labeling for stereo-electroencephalography (sEEG).

Methods: Automated localization was carried out by intensity-upscaling registered post-implantation CT scans followed by volume-based clustering of resulting metal artifacts. For each probe, an initial list of coordinates for each electrode are estimated from the trajectory axis computed from entry and target point coordinates extracted from the implant log, the distance between entry and target point coordinates, and the inter-electrode distance. Probe names are retrieved from the implant log and used to automatically label each electrode. Trajectories are then fit using linear regression modeling to facilitate the identification of metal electrode artifacts from noise. Artifact clusters identified using volumetric clustering search algorithms are iteratively searched while masking out any cortical region not within the current cluster of interest to resolve overlapping or conflated artifacts. Identified clusters are aligned to the robotic trajectories using information about the parallel relationships between the trajectory path and trajectories of the identified clusters as well as information about the centroid of the clusters to refit the trajectories.

Results: Electrode localization accuracy was evaluated for 15,926 electrodes (PMT, DIXI) across 72 patients, and coordinates derived from automatic localization were compared to coordinates obtained from manual localization. All electrodes were labeled correctly using our automated approach without any additional user input. The average localization error across all electrodes was 1.85 +/- 4.79 mm. Using a linear mixed-effects analysis, we found that probe location (F=41.03, p< 0.01), electrode size (F=16.03, p< 0.01), and distance to the target coordinate (F=40.39, p< 0.01) significantly affected the localization error, while the brand of the probe (F=0.27, p=0.61) did not.