Automatic segmentation of stereoelectroencephalography (SEEG) electrodes post-implantation considering bending

Granados, Alejandro; Vakharia, Vejay; Rodionov, Roman; Schweiger, Martin; Vos, Sjoerd B.; O’Keeffe, Aidan G.; Li, Kuo; Wu, Chengyuan; Miserocchi, Anna; McEvoy, Andrew W.; Clarkson, Matthew J.; Duncan, John S.; Sparks, Rachel; Ourselin, Sebastien

doi:http://dx.doi.org/10.1007/s11548-018-1740-8

Automatic segmentation of stereoelectroencephalography (SEEG) electrodes post-implantation considering bending

License

Citation

Granados A, Vakharia V, Rodionov R, Schweiger M, Vos SB, O’Keeffe AG, et al. Automatic segmentation of stereoelectroencephalography (SEEG) electrodes post-implantation considering bending. International Journal of Computer Assisted Radiology and Surgery. 2018 Jun;13(6):935-46. DOI: 10.1007/s11548-018-1740-8

Abstract

The accurate and automatic localisation of SEEG electrodes is crucial for determining the location of epileptic seizure onset. We propose an algorithm for the automatic segmentation of electrode bolts and contacts that accounts for electrode bending in relation to regional brain anatomy. Co-registered post-implantation CT, pre-implantation MRI, and brain parcellation images are used to create regions of interest to automatically segment bolts and contacts. Contact search strategy is based on the direction of the bolt with distance and angle constraints, in addition to post-processing steps that assign remaining contacts and predict contact position. We measured the accuracy of contact position, bolt angle, and anatomical region at the tip of the electrode in 23 post-SEEG cases comprising two different surgical approaches when placing a guiding stylet close to and far from target point. Local and global bending are computed when modelling electrodes as elastic rods. Our approach executed on average in 36.17 s with a sensitivity of 98.81% and a positive predictive value (PPV) of 95.01%. Compared to manual segmentation, the position of contacts had a mean absolute error of 0.38 mm and the mean bolt angle difference of 0.59◦ resulted in a mean displacement error of 0.68 mm at the tip of the electrode. Anatomical regions at the tip of the electrode were in strong concordance with those selected manually by neurosurgeons, ICC(3, k) = 0.76, with average distance between regions of 0.82 mm when in disagreement. Our approach performed equally in two surgical approaches regardless of the amount of electrode bending. We present a method robust to electrode bending that can accurately segment contact positions and bolt orientation. The techniques presented in this paper will allow further characterisation of bending within different brain regions.