SEEG assistant: a 3DSlicer extension to support epilepsy surgery

Abstract

Background: In the evaluation of Stereo-Electroencephalography (SEEG) signals, the physicist's workflow involves several operations, including determining the position of individual electrode contacts in terms of both relationship to grey or white matter and location in specific brain regions. These operations are (i) generally carried out manually by experts with limited computer support, (ii) hugely time consuming, and (iii) often inaccurate, incomplete, and prone to errors.

Results: In this paper we present SEEG Assistant, a set of tools integrated in a single 3DSlicer extension, which aims to assist neurosurgeons in the analysis of post-implant structural data and hence aid the neurophysiologist in the interpretation of SEEG data. SEEG Assistant consists of (i) a module to localize the electrode contact positions using imaging data from a thresholded post-implant CT, (ii) a module to determine the most probable cerebral location of the recorded activity, and (iii) a module to compute the Grey Matter Proximity Index, i.e. the distance of each contact from the cerebral cortex, in order to discriminate between white and grey matter location of contacts. Finally, exploiting 3DSlicer capabilities, SEEG Assistant offers a Graphical User Interface that simplifies the interaction between the user and the tools. SEEG Assistant has been tested on 40 patients segmenting 555 electrodes, and it has been used to identify the neuroanatomical loci and to compute the distance to the nearest cerebral cortex for 9626 contacts. We also performed manual segmentation and compared the results between the proposed tool and gold-standard clinical practice. As a result, the use of SEEG Assistant decreases the post implant processing time by more than 2 orders of magnitude, improves the quality of results and decreases, if not eliminates, errors in post implant processing.

Conclusions: The SEEG Assistant Framework for the first time supports physicists by providing a set of open-source tools for post-implant processing of SEEG data. Furthermore, SEEG Assistant has been integrated into 3D Slicer, a software platform for the analysis and visualization of medical images, overcoming limitations of command-line tools.

Keywords: Automatic segmentation; Epilepsy; Epileptic zone detections; GMPI; Medical imaging; SEEG.

Fig. 1

SEEG Assistant Framework is composed of three interconnected elements. The figure illustrates the three modules constituting the backbone of SEEGA module (light green box), their relations (dark green arrow), and I/O. Contact Position Estimator (CPE) extracts the contact positions from post-implant CT data provided the entry/target points for each implanted electrode (Fiducial File) and several parameters. These latter consist in the implanted electrode model (i.e., number contacts and inter-contact distance) and two boolean parameters. CPE outputs a fiducial file (Recon File) containing all segemented channel positions in the reference image (i.e., CT scanner space) along with contact label. In the test settings, electrodes are marked with a single capital letter (e.g., A or B) and channels are identified with increasing numbers (e.g., A1). The next two modules namely Grey Matter Proximity Index Estimator (GMPIE) and Brain Zone Detector (BZD) can be used without a specific order since they do not depends on each other. The former uses the segmented channel positions and pial/white matter Surfaces (e.g. Freesurfer) and computes for each contact the distance from white/grey boundary normalized to the cortical thickness. The latter uses contact positions and a volumetric probabilitic parcellation (e.g. Destrieux) and estimates the probability (i.e., proximity) of given source/anatomical area to be the generator of the recorded electrophysiological activity. All these information are added to the Recon File that can be saved for later usage or inspected from 3DSlicer interface

Fig. 2

Interfaces and workflow of SEEGA. This figure shows the interfaces designed to interact with the underlying algorithms and provides an example of the workflow. All drop-down listboxes are populated with data included in the 3DSlicer scene. a Initial configuration panel of CPE where a fiducial file needs to be selected. b Shows the path of the deetoS binary file provided within SEEGA installation. Once a fiducial file (seeg) has been selected a table is populated with recognized electrodes with valid entry/target pairs. The drop-down list next to the electrode label is used to select its specific model which univocally defines number of contacts and inter-contact distances. An example of what is shown in the 3D View panel wihtin 3DSlicer is provided as inset with post-implant thresholded-CT meshes (ct-post - red) and fiducial file shown as markups (black dots) with letters representing each electrode. c Shows the BZD interface where both BZD takes a volumetric parcellation (aparc+aseg - purple star) and the recontructed data (recon). Examples of such inputs are shown below the interface. d Shows GMPIE interface where the five inputs are defined as left (red) right (green) pial (star) and white (pentagon) surfaces together with segmented channel positions (recon)

Fig. 3

CPE out performs manual segmentation in complex and critical cases. a As an example of SEEG complexity, we show MRI and thresholded post-implant CT scans for one subject from our cohort. Contacts are shown as groups of white voxels. This case illustrates the complexity of SEEG implants with electrode shafts following non-planar directions (e.g. X), shafts targeting almost the same geometrical point (e.g. R and R’). b CPE segments all contacts (green spheres) belonging to each electrode from post-implant CT scans, represented here as red 3D meshes obtained tesselleting the thresholded data to ease visualization. c Show the right pial surface with 3D post-implant thresholded-CT meshes and the cut plane used in panel d where the example of X and X’ electrodes are shown. Those examples represent the case of non-planar insertion trajectories which yielded an artefactually fused electrode. CPE integrating the knowledge of the electrode model can segment the contact positions more accurately than visual inspection

Fig. 4

CPE module provides more accurate results compared to manual segmentation. a Contact distances from axis are on average similar between manual (blue) and automatic (red) segmentation. In general automatic segmentation performs better in keeping all contacts aligned to their axis. Probability distribution shows that the paired differences of automatical and manual defined contact to their axis is signficantly (p < 0.05, paired wilcoxon test) smaller than zero. As expected, manual segmentation yields higher variability and a larger fraction of outliers. b Inter-Contact distributions show gaussian-like distributions. Probability distribution of paired difference of automatic and manually defined contacts shows that our method significantly (p < 0.05) out performs compared to the manual case in representing real inter-contact distance. Finally, automatic segmentation shows a smaller variability compared to manual segmentation

Fig. 5

BZD estimates most probable electrical sources using volumetric atlases. a Number of contacts recording from each parcel is shown on top of the inflated surfaces of the cerebral cortex. Color code is shown in the color bar and represents the less (white) to most (red) sampled regions. Two atlases have been used to test the algorithm: Desikan-killeany (parc68 - left column) and Destrieux (parc2009 - right column). Both atlases yields a similar spatial distribution. The existing differences can be due to the different parcel resolutions. b Number of contacts from each parcel for both parc68 and parc2009 are shown as color bar histograms divided between cerebral lobes. Here similar patterns can be seen across atlases

Fig. 6

GMPI reflects channel position relative to cerebral cortex. a An MRI inset is shown with atlas (i.e., Destrieux) and three meshes are shown representing Hippocampus (green) Amygdala (yellow) and pial (red) surfaces. Segmented contact positions are represented as pink spheres centered in the estimated position. b This panel shows a zoom-in on the segmented contact plane where it can be seen the contact positions, the intersection between plane and subcortical structures (i.e., Hip and Amy). Cortical sheet has been marked with brown color. The axis below shows that GMPI decreases while the shaft penetrates white matter fibers with increasing distance from cortical sheet. c Probability (top) and cumulative (bottom) distributions of GMPI values across all cortical contacts