Poster K6 - Comparison of Thalamic Atlases and Segmentation Techniques in Defining Motor Nuclei for Deep Brain Stimulation Targeting in Essential Tremor
Assistant Professor of Biomedical Engineering Bucknell University Lewisburg, Pennsylvania, United States
Introduction: Deep brain stimulation (DBS), a surgical procedure using implanted electrodes to stimulate specific brain regions, is a proven treatment for the involuntary trembling symptoms of essential tremor (ET). Research demonstrates that DBS yields a long-term reduction of tremor while also being reversible. The thalamus, a nucleus of the basal ganglia, is an established stimulation target. It can be divided into smaller subnuclei, with the ventral intermediate (VIM) nucleus being the main target for DBS. When stimulating motor subnuclei like the VIM, DBS yields therapeutic results.
Since clinical MRI does not have the resolution to allow for observable partitioning of thalamic subnuclei, neurosurgeons rely on brain atlases (anatomical maps) to locate relevant motor subnuclei for stimulation. However, there are many thalamic atlases with widely disparate delineations of subnuclei involved in motor and sensory signaling. This variety comes from individual topographical differences, the influence of age and disease, and the different weights placed on cytoarchitectural anatomy and functional connectivity when segmenting the thalamus.
In addition to inter-atlas variability, traditional approaches to thalamic segmentation do not consider variability across patients. In such approaches, the atlas acts as a stencil, where its borders are used to segment a patient’s thalamus. In more data-driven approaches, diffusion tensor imaging (DTI) is used to form clusters of voxels (3-D pixels) exhibiting similar tissue properties.
The objective of this research was to compare the effectiveness of multiple atlases and segmentation techniques in defining thalamic motor and sensory regions by analyzing the resulting segmentations’ correspondence with clinical outcomes.
Materials and
Methods: 22 ET patients who underwent unilateral VIM DBS were analyzed retrospectively. Each patient had an MRI-derived thalamus, the thalamus’ corresponding DTI data, and volumes of tissue activation (VTA) associated with clinical outcomes: tremor reduction (22 VTAs) and paresthesia, a sustained tingling sensation (32 VTAs).
Six prominent thalamic atlases with different motor-sensory boundaries and subnuclei counts were obtained from Lead-DBS, a toolbox allowing for DBS electrode reconstruction using patient-specific imaging. Each atlas was registered to each patient’s thalamus using an iterative closest point algorithm (Fig. 1).
Two segmentation approaches were used: in atlas-based segmentation, patient thalamus voxels were segmented based on the atlas subnuclei they resided in or were closest to. In DTI-based segmentation, the atlas subnuclei centroids were used as seed points for a k-means clustering algorithm that formed subnuclei clusters based on DTI-derived tissue properties. The stability of the DTI-based approach was assessed by randomly shifting the seed points, clustering using k-means, and measuring the consistency of the results using the Dice coefficient (DC=1 indicates perfect overlap; DC=0 indicates zero overlap).
The resulting anatomy of the motor and sensory regions from the two thalamic segmentation approaches was analyzed by comparing volume, DC, surface area, and centroid position. The overlap between VTAs and motor and sensory regions was calculated and compared with the associated clinical outcome (tremor reduction and paresthesia) between segmentation approaches. These analyses were done across atlases. Paired, two-sided Wilcoxon signed-rank tests were used to determine significance. All parts of the analysis were performed in MATLAB.
Results, Conclusions, and Discussions: The stability of the DTI-based clustering algorithm varied across atlases. The most stable atlas was Jakab (DC=0.82±0.09) and the least stable atlas was Iglesias (DC=0.58±0.26) when comparing the motor and sensory regions resulting from unshifted and shifted seeding. The volume of motor and sensory regions varied between segmentation approaches and across atlases (Fig. 2). Overall, atlas-based segmentation produced a significantly larger motor volume (p=3.87E-9) and smaller sensory volume (p=9.19E-8) than DTI-based segmentation (aggregating across all patients and atlases).
For stimulation associated with tremor reduction, the atlas-based approach showed significantly greater VTA-motor overlap than VTA-sensory overlap in 6/6 atlases; 3/6 atlases for the DTI-based approach showed significance (Fig. 3). Moreover, VTA-sensory overlap was greater for paresthesia VTAs than for tremor reduction VTAs in 5/6 atlases when using atlas-based segmentation. In contrast, DTI-based segmentation yielded similar results in 3/6 atlases and opposite results in 2/6 atlases (Fig. 4).
Atlases with more subnuclei had stabler clustering results because the increased number of final clusters, when aggregated, reduced the effect of shifted initial seeding. Non-zero motor activation and minimal sensory activation were expected for therapeutic stimulation, and atlas-based segmentation yielded this trend across all atlases. In general, atlas-based segmentation produced larger motor volume and smaller sensory volume compared to DTI-based segmentation, which may explain the correspondence to the expected clinical trend.
Increased VTA-sensory overlap was expected for paresthesia VTAs compared to tremor reduction VTAs, and atlas-based segmentation similarly yielded this trend across all atlases. It is important to note that the k-means clustering algorithm weighted tissue properties and physical distance equally by default; modifying these weights could enhance individual atlas performance.
Atlas-based segmentation was found to be a more generalizable thalamic segmentation method than DTI-based segmentation for use across atlases. Furthermore, the resulting clusters from the atlas-based approach aligned more closely with clinical outcomes. The Jakab atlas exhibited alignment with clinical outcomes in both atlas-based and DTI-based segmentation and had the stablest segmentation results, making it the most clinically accurate and robust atlas that was investigated.
Acknowledgements (Optional): The authors would like to thank the patients for their participation in this study, our collaborators at the University of Michigan for the clinical data, and the Lead-DBS group for providing the thalamic atlases and extending support. This work was supported by the Costa Healthcare Research & Design Fund at Bucknell University.
