Anatomy-driven multiple trajectory planning (ADMTP) of intracranial electrodes for epilepsy surgery

One-third of patients with focal epilepsy have poor seizure control despite pharmaceutical treatment [11]. These patients are candidates for curative resective surgery if the epileptogenic zone (EZ) can be located and the EZ does not overlap with eloquent cortex. Non-invasive imaging can definitively identify the EZ in many patients; however, approximately 25% require invasive surgical intervention to identify the EZ [5]. One such intervention involves implanting stereo-electroencephalographic (SEEG) intracranial depth electrodes to target deep and superficial regions of interest (ROIs) within the cortex. Implanted electrodes may also map eloquent areas (e.g. motor or sensory cortex) and aid in determining safe resection margins [6].

Implantation planning of SEEG electrodes is a two-stage process. In the first stage, a strategy consisting of a list of deep and superficial ROIs to record from is generated by a multi-disciplinary team of epileptologists, neurophysiologists, and neurosurgeons. Figure 1a displays an example strategy for a patient. In the second stage, neurosurgeons use the list of ROIs to guide precise planning of each electrode trajectory. Figure 1b displays the implantation plan with the ROI each electrode must attain. Figure 1c displays the implantation plan with blood vessels, important to consider when placing trajectories to minimize the risk of haemorrhage, and the skull, important to ensure the skull entry is surgically feasible. Finally, electrode trajectories are also placed to maximize their recording from grey matter (GM), as epileptic seizures arise in GM rather than white matter.

In current clinical practice, electrode trajectories are determined by manual placement and visual inspection of preoperative imaging data to ensure all surgical criteria are satisfied [13]. This is a complex, time-consuming task typically requiring between 2 and 3 h per patient. Automated trajectory planning algorithms may reduce planning time, improve safety, and increase GM sampling by optimizing trajectories according to quantitative measures of suitability.

Automated single trajectory planning algorithms have been presented to maximize the distance to blood vessels and satisfy other surgical constraints for individual trajectories [2, 7, 8, 14, 17, 18, 20]. Each method defines a unique risk score based on specific surgical criteria, in general, trajectories near blood vessels or other critical structures are given high weights, and the lowest risk trajectory is selected. Most of these methods require the user to specify a target as a point [2, 8, 14, 17, 18, 20]. Essert et al. [7] requires the user to manually specify an ROI, as this method was applied to deep brain stimulation (DBS) the ROI was either the subthalamic nucleus or the ventral intermediate nucleus. This method also constrained the entry point search to a user-defined region of the skull, greatly reducing the potential number of trajectories. These methods are limited when applied to SEEG electrode implantation as (1) the most suitable target point for deep ROIs may not be obvious or may correspond to large brain structures, (2) multiple electrodes are implanted and must be placed to avoid conflicts between electrodes and to maximize GM sampling.

Automated multiple trajectory planning algorithms have been designed for implantation planning of SEEG electrodes. These methods allow users to specify multiple target and entry ROIs [4, 16, 19]. de Momi et al. [4] determined trajectories by randomly sampling from user-defined target and entry ROIs. Each ROI was defined as a user-selected point and distance threshold; hence, each target can be considered a spherical ROI. The best trajectory for each electrode was found in terms of entry angle with the skull, avoidance of blood vessels, avoidance of sulci, and ensuring a minimum distance between electrodes. Sparks et al. [16] required the user to specify each target as a point, and entry points were defined by the points on the skull surface below an angle threshold. The best trajectory for each electrode was then found in terms of avoidance of critical structures, maximum GM sampling, and ensuring a minimum distance between electrodes. As in the case of single trajectory planning algorithms, target points may not be obvious.

Zelmann et al. [19] overcame the need to define candidate target points. The hippocampus and amygdala were selected as ROIs and targets were computed by sampling a Gaussian distribution defined on the ROI distance map. Hence, targets near the ROI centreline were preferentially sampled. The best trajectory for each electrode was found in terms of entry angle with the skull, avoidance of critical structures, maximum GM sampling, and ensuring a minimum distance between electrodes. The major limitations of this work are that only hippocampus and amygdala were used as ROIs. Furthermore, the assumption that points near the ROI centreline are desirable may not hold for all ROIs.

In Sparks et al. [15], we presented anatomy-driven multiple trajectory planning (ADMTP) to compute electrode trajectories from a list of user-defined ROIs. This method is flexible enough to target any anatomically defined ROI. However, in the previous ADMTP implementation it was assumed that the centre of the ROI was the optimal target. In this manuscript, we have improved the ADMTP algorithm [15] in the following respects:Additionally, we have performed a thorough qualitative analysis comparing ADMTP and manual implantation plans according to safety, surgical feasibility, and spatial coverage of the suspected EZ for trajectories, target and entry points.

MR imaging is performed on a GE 3T MR750 scanner with a 32-channel head coil. A 3D T1-weighted MPRAGE scan is performed with a field-of-view (FOV) of \(224 \times 256 \times 256\) mm (\(\hbox {AP} \times \hbox {LR} \times \hbox {IS}\)) with an acquisition matrix of \(224 \times 256 \times 256\) for a reconstructed voxel size of 1 mm isotropic (TE/TR/TI \(=\) 3.1 / 7.4 / 400 ms; flip angle \(11^{\circ }\); parallel imaging acceleration factor 2). A post-gadolinium T1-weighted scan is performed with an FSPGR sequence with a FOV of \(224 \times 256 \times 256\) mm and acquisition and reconstruction matrix of \(224 \times 256 \times 256\) (TE/TR \(=\) 3.1 / 7.4 ms; flip angle \(11^{\circ }\)). MR angiography (MRA) and venography (MRV) are performed using a 3D phase-contrast sequence with an FOV of \(220 \times 220 \times 148.8\) mm with an acquisition matrix of \(384 \times 256 \times 124\) for a reconstructed voxel size of \(0.43 \times 0.43 \times 0.6\) mm (flip angle \(8^{\circ }\); parallel imaging acceleration factor 2). To highlight the arteries, the MRA is scanned with a velocity encoding of 80 cm/s (TE/TR \(=\) 4.0 / 9.3 ms). For sensitivity to the venous circulation, the MRV is scanned with a velocity encoding of 15 cm/s (TE/TR \(=\) 4.8 / 26.4 ms), fat suppression, and a saturation band inferior to the FOV.

The algorithms chosen for brain parcellation and segmentation of GM, arteries, veins, sulci, and skull are currently part of routine clinical practice for manual planning of SEEG electrode trajectories [12, 13]. Figure 2 shows examples of brain parcellation and structure segmentation. Veins and arteries are segmented from a post-gadolinium T1-weighted scan, MRV, and MRA using a multi-scale, multi-modal tensor voting algorithm [21]. The skull is segmented from CT using thresholding followed by morphologic operations for a connected surface.

Geodesic information flows (GIF) [3] is used to perform brain parcellation on a 3D T1-weighted MPRAGE using the Brain Collaborative Open Labeling Online Resource (BrainCOLOR) atlas [10] to define the anatomical labels. The brain atlas contains 142 possible regions. GM and sulci are extracted from the brain parcellation using thresholding.

Candidate target point selection is performed as follows. A risk map of the ROI is calculated as described in “Target risk map computation” section (as in [15]). Next, a novel algorithm is used to determine M unique target points \(T_{n,i} : i \in \{1, \ldots , M\}\) for each electrode in the ROI, which are computed using K-means clustering of the candidate target points as described in “Candidate target point sampling” section.

A set of possible entry points defined as \(E_{n,j} : j \in \{1, \ldots , P\}\) are computed by considering all vertices on the skull mesh. Potential trajectories \(\overline{T_{n,i}E_{n,j}}\) are removed from consideration with the following criteria (modified from Zombori et al. [20]):If no suitable trajectories are found, the \(d_\mathrm{ang}\) and \(d_\mathrm{len}\) are iteratively relaxed such that \(d_\mathrm{len}\) is increased by 10 mm (up to 110 mm) and \(d_\mathrm{ang}\) is increased by \(10^{\circ }\) (up to \(45^{\circ }\)) until a set of suitable trajectories is found.

For suitable trajectories, a weighted score \(S_{n,i,j}\) representing a combination of a risk score \(R_{n,i,j}\) and GM ratio \(G_{n,i,j}\) is calculated. \(R_{n,i,j}\), a measure of cumulative distance to blood vessels, is computed as:where trajectories with \(f_\mathrm{cri}(x)\) closer than \(d_\mathrm{safe}\) have the highest risk (\(R_{n,i,j}=1\)) while \(f_\mathrm{cri}(x)\) farther than \(d_\mathrm{risk}\) have no risk (\(R_{n,i,j}=0\)).

\(G_{n,i,j}\) is a measure of the proportion of electrode contacts, the parts of the electrode that record EEG signals, in GM. The electrode contact locations were modelled by selecting a commonly implanted electrode. Each electrode has \(Q=10\) contacts placed at the points \(p_q\) spaced in intervals of 10 mm along the trajectory starting at the target point \(T_{n,i}\). Each contact is assessed on location in the GM at three points, the centre of the contact, \(p_{q}\), and both ends of the contact, \(p_q \pm p_r\), where \(p_{r} = 1.2\) mm is the radius of the contact. Hence, each electrode contact may have a value of \(\big [0, 1/3, 2/3, 1\big ]\). \(G_{n,i,j}\) is the summation of GM capture over all contacts calculated as,where \(f_\mathrm{gm}(\cdot )\) is the signed distance from the GM surface and \(H[\cdot ]\) is the Heaviside function, with values of 1 inside GM and 0 outside.

The weighted score is computed as \(S_{n,i,j} = 10 * R_{n,i,j} + G_{n,i,j}\). 10 was set empirically to prioritize low risk over a high GM ratio.

The final implantation plan V(N) is found by optimizing,where \(d_\mathrm{traj}\) specifies the minimum distance between trajectories to ensure no conflicts. V(N) is computed using the depth-first graph search method described in Sparks et al. [16].

One modification has been made to the graph search strategy to improve performance. For an electrode conflict \(D(\overline{T_{n,i} E_{n,j}}, \overline{T_{k,i} E_{k,j}} ) < d_\mathrm{traj}\) let the next considered trajectory be defined as \(\overline{T_{k',i} E_{k',j}}\). In Sparks et al. [16], \(\overline{T_{k',i} E_{k',j}}\) was the next lowest risk trajectory (i. e. \(k+1\)). To decrease computation time in this work, we calculate \(\overline{T_{k',i} E_{k',j}}\) as the next lowest risk trajectory that satisfies the constraint \(D(\overline{T_{k,i} E_{k,j}}, \overline{T_{k',i} E_{k',j}}) > 0.5 \text { mm}\). This allows for many spatial similar trajectories to be discarded decreasing the number of trajectory combinations considered. If no combination of trajectories exists which satisfies \(d_\mathrm{traj}\), ADMTP returns the plan with the largest possible distance between trajectories.

ADMTP was evaluated on retrospective data from 20 patients with refractory focal epilepsy who underwent intracranial electrode implantation. Each patient had between 7 and 12 electrodes placed (190 total). Prior to trajectory planning a list of regions to target is determined by a multi-disciplinary team discussion consisting of expert epileptologists, neurophysiologists and neurosurgeons based on the evaluation of the clinical history, seizure semiology, ictal and interictal scalp EEG and results of multi-modal imaging. Following this manual plans (MPs) were determined by the consensus of two neurosurgeons, and some regions were removed from consideration if no safe trajectory to the target could be determined. The list of regions where a MP trajectory was determined were used to evaluate ADMTP with the parameters given in Table 1. ADMTP and MP trajectories corresponding to the same region were compared.

All 190 trajectories were assessed on the following quantitative measures of trajectory suitability: angle with respect to the skull surface normal, risk score (\(R_{n,i,j}\)), distance to nearest critical structure, and GM ratio (\(G_{n,i,j}\)). In Fig. 4, each point corresponds to one electrode with the manual trajectory value plotted on the X axis and the ADMTP trajectory value plotted on the Y axis. The red point represents the centre of mass for each measure. Points below the diagonal have a lower value for ADMTP compared to manual trajectories. For angle and risk score, points below the diagonal correspond to ADMTP giving the preferred value. For GM ratio and minimum distance points above the diagonal represent ADMTP giving the preferred value. A two-tailed Student’s t test evaluated the statistical significance between values determined by ADMTP and manual trajectories where the null hypothesis was that the methods return similar values.

ADMTP had a more feasible entry angle in 129 / 190 trajectories and increased GM sampling in 108 / 190 trajectories. ADMTP found trajectories that were safer in terms of reduced risk score (159 / 190 trajectories) and increased distance to the closest critical structure (155 / 190 trajectories). All differences between ADMTP and manual trajectories were statistically significant. The discrepancies between risk score and closest critical structure in 4 trajectories are due to risk being a cumulative distance measure; hence, it is maybe higher due to other critical structures that are not the closest critical structure.

A neurosurgeon who did not participate in creating MPs and was blinded to plan origin assessed 10 randomly selected MPs and corresponding ADMTPs (94 electrodes). Implantation plans as a whole were assessed for suitable spatial coverage of the suspected EZ. Each trajectory was assessed according to whether they would proceed to implantation based on the following criteria.In all cases, safety was assessed by being at least 3 mm away from all visible blood vessels and not crossing sulcal boundaries. In this study, crossing sulcal boundaries was included in safety criteria as small vessels within deep sulci may not be fully resolved with the current imaging protocol.

All implantation plans had suitable spatial coverage of the EZ. Figure 5 shows the results of the qualitative assessment per plan. MPs attained the appropriate target in 97% of cases, while ADMTP attained the appropriate target in 90% of cases. Table 2 details the reasons why target points were considered unsuitable, primarily due to placement in the wrong portion of the ROI. Table 3 lists the reasons why entry points and trajectories were considered infeasible. In both MPs and ADMTP plans, 30% of trajectories were considered too close to either blood vessels or sulci. If proximity to sulci was not included in the safety assessment, 85% of MP and 95% of ADMTP trajectories were considered suitable.

The computational efficiency of ADMTP was assessed for target point selection, trajectory risk scoring, and implantation plan computation. Computations were performed on a computer with a Intel(R) Xeon(R) 12 core CPU 2.10 GHz with 64.0 GB RAM and a single NVIDIA Quadro K4000 4GB GPU. Figure 6 reports computation time for each step in the process.

Automated planning took 54.5 s (17.3–191.9 s). For the majority of plans, most of the computation time was spent on trajectory risk scoring (11–146 s). As described in Zombori et al. [20] risk score computation time is dependent on the number of trajectories being considered. For target point search computation time is relatively linear with respect to the number of electrodes. Finally, for cases with a fewer number of electrodes (<10) plan computation time was insignificant (under 1 s), but this computation time increases substantially for more complex plans with a larger number of electrodes.