Elsevier

NeuroImage

Volume 52, Issue 1, 1 August 2010, Pages 69-85
NeuroImage

Reproducibility of thalamic segmentation based on probabilistic tractography

https://doi.org/10.1016/j.neuroimage.2010.04.024Get rights and content

Abstract

Reliable identification of thalamic nuclei is required to improve targeting of electrodes used in Deep Brain Stimulation (DBS), and for exploring the role of thalamus in health and disease. A previously described method using probabilistic tractography to segment the thalamus based on connections to cortical target regions was implemented. Both within- and between-subject reproducibility were quantitatively assessed by the overlap of the resulting segmentations; the effect of two different numbers of target regions (6 and 31) on reproducibility of the segmentation results was also investigated. Very high reproducibility was observed when a single dataset was processed multiple times using different starting conditions. Thalamic segmentation was also very reproducible when multiple datasets from the same subject were processed using six cortical target regions. Within-subject reproducibility was reduced when the number of target regions was increased, particularly in medial and posterior regions of the thalamus. A large degree of overlap in segmentation results from different subjects was obtained, particularly in thalamic regions classified as connecting to frontal, parietal, temporal and pre-central cortical target regions.

Introduction

Deep brain stimulation (DBS) of thalamic nuclei is being developed as a treatment for drug-resistant epilepsy (Theodore and Fisher, 2004) and is widely used in the treatment of movement disorders (Koller et al., 1997). Reliable identification of thalamic nuclei has the potential to improve treatment success, by improving the accuracy with which DBS electrodes are targeted. Currently, targeting of DBS electrodes uses qualitative magnetic resonance images in combination with thalamic atlases, for example, the Morel thalamic atlas (Morel et al., 1997) or Schaltenbrand and Wahren atlas (Schaltenbrand and Wahren, 1977). Typically, the atlases used are based on a small number of subjects and do not take into account inter-subject variability. Furthermore, atlases may present data from different subjects sectioned in three orthogonal planes, and there is good evidence that these orthogonal sections do not correspond sufficiently to provide reliable targeting (Nowinski et al., 2008). New methods which reliably and accurately identify individual nuclei within the thalamus are required to improve targeting in DBS, and will also have great potential for examination of the integrity of thalamic nuclei across a range of diseases.

Recent work has shown that thalamic sub-regions, thought to correspond to groups of thalamic nuclei, can be identified from diffusion imaging data at the single-subject level using connectivity-based segmentation (Behrens et al., 2003a, Johansen-Berg et al., 2005). Each thalamus contains different nuclei that can be distinguished histologically by their characteristic cyto- and myelo-architecture (Morel et al., 1997). Currently, however, these nuclei cannot be identified on conventional, qualitative MR images due to insufficient contrast in deep grey matter structures. The different nuclei are known to have extensive connections to specific regions of the ipsilateral cerebral cortex (as well as reciprocal connections to each other). Using a probabilistic tractography method, Behrens et al. (2003a) segmented the thalamus based on connections between thalamic voxels and ipsilateral cortical target regions but carried out only a qualitative assessment of the between-subject reproducibility of the method. Segmentation of the thalamus was carried out for eight subjects using four large cortical target regions. For six of the eight subjects, the resulting segmentations were found to be very similar. The diffusion data for the remaining two subjects were of lower quality and hence produced a poorer segmentation. No quantitative assessment was carried out of the reproducibility of the method for a single subject scanned multiple times. The results have been promising, producing thalamic segmentations which appear to be reproducible across subjects (Behrens et al., 2003a, Johansen-Berg et al., 2005), but detailed quantitative assessment of within- and between-subjects reliability has not yet been attempted.

In the study reported here, a quantitative assessment of the reproducibility of the connectivity-based method of thalamic segmentation was carried out, both within and between subjects. Initially, we examined reproducibility of the thalamic segmentation method in a single subject. We approached this problem using multiple re-analyses of the same single dataset, varying the starting conditions of the probabilistic tractography algorithm. This was followed by an examination of multiple datasets that had been independently acquired from the same subject. Subsequently we examined the reproducibility of thalamic segmentation between subjects in a group of normal subjects. Additionally, we investigated whether reproducibility of our findings was a function of the number of cortical targets, and whether the degree of contrast in the structural imaging data influenced the definition of cortical and thalamic regions.

Section snippets

Data acquisition

Sixteen DTI datasets were acquired for a single healthy subject (male, 24 years). These data were acquired in 12 sessions over a period of eight months. In four of the sessions, two DTI datasets were obtained (datasets 1 and 2, 3 and 4, 5 and 6, 15 and 16). The remaining eight datasets were acquired in separate sessions. A 3T, GE Signa HDx system (General Electric, Waukesha, WI, USA), with actively shielded magnetic field gradients and a maximum gradient amplitude of 40 mT m1, was used. The

Results

Cortical and sub-cortical regions specific to the first subject were generated based on the IR-FSPGR image and the inverted-contrast T1 map. Fig. 1 shows an axial slice from each of these structural images, along with the corresponding cortical and sub-cortical segmentations derived from these images. Cortical regions defined from both structural images were similar; TAO between the two sets of results was 0.753. Differences were observed in the thalamus masks generated from the two images: for

Discussion

This work demonstrates that using connectivity-based analysis to segment the thalamus on MR images of the human brain yields subdivisions with high reproducibility. Although the method relies on repeated sampling of a Probability Density Function (PDF), we showed that segmentations produced by reanalysing the same dataset using different starting conditions were extremely similar. Additionally, we showed within-subject reliability of the method with stable segmentation results obtained from

Acknowledgments

MPR is supported by the Getty Family Foundation. WRC acknowledges financial support from the Department of Health via the National Institute for Health Research (NIHR) specialist Biomedical Research Centre in Mental Health award to South London and Maudsley NHS Foundation Trust. RAH was supported by a grant from the Dunhill Medical Trust.

References (29)

  • J.L.R. Andersson et al.

    Non-linear optimisation

    (2007)
  • T.E. Behrens et al.

    Non-invasive mapping of connections between human thalamus and cortex using diffusion imaging

    Nat. Neurosci.

    (2003)
  • T.E. Behrens et al.

    Characterization and propagation of uncertainty in diffusion-weighted MR imaging

    Magn. Reson. Med.

    (2003)
  • W.R. Crum et al.

    Generalized overlap measures for evaluation and validation in medical image analysis

    IEEE Trans. Med. Imaging

    (2006)
  • Cited by (0)

    View full text