Automatic skull segmentation from MR images for realistic volume conductor models of the head: Assessment of the state-of-the-art

Highlights

• Assessment of three methods for the automatic skull segmentation from MR images.

• Rigorous test of their accuracy by comparison against CT data of the same subjects.

• FSL and SPM12 can achieve reasonable accuracy for the upper part of the head.

• A combination of T1- and T2-weighted images, rather than a single T1, is suggested.

• Accuracy strongly benefits from optimization of the MRI sequence parameters.

Abstract

Anatomically realistic volume conductor models of the human head are important for accurate forward modeling of the electric field during transcranial brain stimulation (TBS), electro- (EEG) and magnetoencephalography (MEG). In particular, the skull compartment exerts a strong influence on the field distribution due to its low conductivity, suggesting the need to represent its geometry accurately. However, automatic skull reconstruction from structural magnetic resonance (MR) images is difficult, as compact bone has a very low signal in magnetic resonance imaging (MRI). Here, we evaluate three methods for skull segmentation, namely FSL BET2, the unified segmentation routine of SPM12 with extended spatial tissue priors, and the skullfinder tool of BrainSuite. To our knowledge, this study is the first to rigorously assess the accuracy of these state-of-the-art tools by comparison with CT-based skull segmentations on a group of ten subjects. We demonstrate several key factors that improve the segmentation quality, including the use of multi-contrast MRI data, the optimization of the MR sequences and the adaptation of the parameters of the segmentation methods. We conclude that FSL and SPM12 achieve better skull segmentations than BrainSuite. The former methods obtain reasonable results for the upper part of the skull when a combination of T1- and T2-weighted images is used as input. The SPM12-based results can be improved slightly further by means of simple morphological operations to fix local defects. In contrast to FSL BET2, the SPM12-based segmentation with extended spatial tissue priors and the BrainSuite-based segmentation provide coarse reconstructions of the vertebrae, enabling the construction of volume conductor models that include the neck. We exemplarily demonstrate that the extended models enable a more accurate estimation of the electric field distribution during transcranial direct current stimulation (tDCS) for montages that involve extraencephalic electrodes. The methods provided by FSL and SPM12 are integrated into pipelines for the automatic generation of realistic head models based on tetrahedral meshes, which are distributed as part of the open-source software package SimNIBS for field calculations for transcranial brain stimulation.

Introduction

Volume conductor models of the head are key components of several neuroscientific methods such as electric field simulations for transcranial brain stimulation (TBS) and source localization in electro- (EEG) and magnetoencephalography (MEG). The anatomical accuracy of the head models has a strong influence on the accuracy of the calculated field distributions (Cho et al., 2015; Dannhauer et al., 2011; Eichelbaum et al., 2014; Lanfer et al., 2012; Montes-Restrepo et al., 2014; Wolters et al., 2006) and attempts to use individualized models based on structural magnetic resonance (MR) images are gaining momentum (Vorwerk et al., 2014). Recently available open-source software, including FSL (Smith et al., 2004), BrainSuite (Shattuck and Leahy, 2002), and SPM12 (http://www.fil.ion.ucl.ac.uk/spm/), facilitates the adoption of this approach by offering automatic segmentation procedures for the head. These tools have been integrated into software pipelines for the forward modeling of electric fields for TBS (e.g., SimNIBS; Thielscher et al., 2015) and EEG/MEG (e.g., FieldTrip; Oostenveld et al., 2011 and Brainstorm; Tadel et al., 2011). Accurate modeling of the skull compartment is an important aspect of individualized head models as the skull strongly shapes the forward solution due to its low conductivity (Dannhauer et al., 2011; Indahlastari et al., 2016; Lanfer et al., 2012; Montes-Restrepo et al., 2014; Stenroos et al., 2014). However, its automatic segmentation is still a major challenge, as the compact bone parts have a very low signal in conventional magnetic resonance imaging (MRI) sequences.

While the performance of most software packages in segmenting the brain have been thoroughly validated, similar tests are scarce for the skull. Thus, in this study we investigate the performance of three widely used neuroimaging software packages, FSL (Smith et al., 2004), BrainSuite (Shattuck and Leahy, 2002), and SPM12 (http://www.fil.ion.ucl.ac.uk/spm/). Specifically, we assess FSL BET2 which includes the BET and betsurf tools (Pechaud et al., 2006), BrainSuite skullfinder (Dogdas et al., 2005), and the unified segmentation routine (Ashburner and Friston, 2005) implemented in SPM12. The latter was tested with spatially extended tissue priors in order to avoid clipping of the lower parts of the head (Huang et al., 2013). In contrast to BrainSuite, FSL and SPM12 support the use of multi-spectral MRI for segmentation. We therefore also compare the results when basing the segmentations on a single, high-resolution T1-weighted structural MR image, as often acquired in neuroimaging studies and used in clinical standard of care, versus a combination of high-resolution T1- and T2-weighted MR images. In addition, for the SPM12-based segmentations, we assess to which extent the results can be improved when applying morphological operations to “clean up” the raw segmentations. We test the quality of the segmentations by systematic comparisons against skull segmentations from computed tomography (CT) scans of the same subjects. To the best of our knowledge, this study is the first to rigorously assess the performance of these tools on skull segmentation and thus serves as important evaluation of the state-of-the-art on this topic.

Whereas the main focus of the paper is on skull segmentation, we further compare the accuracy of the reconstructed brain surfaces derived from SPM12-based segmentations with surfaces obtained using FreeSurfer 5.3.0 (Dale et al., 1999; Fischl et al., 1999). Finally, we exemplarily demonstrate the importance of selecting adequate MRI sequence parameters and adjusting the parameters of the SPM12 segmentation routine to the properties of the MR images in order to achieve robust and accurate results, particularly in non-brain regions. As such, our study gives useful guidelines for the adoption of individualized volume conductor models in neuroscientific research.