Parcellation of the human thalamus based on cortical connectivity information inferred from non-invasive diffusion-weighted images identifies sub-regions that we have proposed correspond to nuclei. Here we test the functional and anatomical validity of this proposal by comparing data from diffusion tractography, cytoarchitecture and functional imaging. We acquired diffusion imaging data in eleven healthy subjects and performed probabilistic tractography from voxels within the thalamus. Cortical connectivity information was used to divide the thalamus into sub-regions with highest probability of connectivity to distinct cortical areas. The relative volumes of these connectivity-defined sub-regions correlate well with volumetric predictions based on a histological atlas. Previously reported centres of functional activation within the thalamus during motor or executive tasks co-localize within atlas regions showing high probabilities of connection to motor or prefrontal cortices, respectively. This work provides a powerful validation of quantitative grey matter segmentation using diffusion tractography in humans. Co-registering thalamic sub-regions from 11 healthy individuals characterizes inter-individual variation in segmentation and results in a population-based atlas of the human thalamus that can be used to assign likely anatomical labels to thalamic locations in standard brain space. This provides a tool for specific localization of functional activations or lesions to putative thalamic nuclei.
The ability to parcellate the human thalamus into meaningful subdivisions in vivo would have significant implications for the study of normal brain function and for investigation of disorders associated with thalamic pathology. Activation of the thalamus has been observed in a number of functional imaging studies involving various sensory (Davis et al., 1998), motor (Samuel et al., 1997; Krams et al., 1998) and cognitive (Baker et al., 1996; LaBar et al., 1999; Dove et al., 2000) tasks, but lack of detailed anatomical information in the thalamus has limited structure–function correlations. Lesions of the thalamus can have wide-ranging effects on sensory, executive and memory functions (Wallesch et al., 1983; von Cramon et al., 1985; Van der Werf et al., 2003) and involvement of thalamo-cortical circuitry is thought to be crucial in certain neurological and psychiatric disorders (Modell et al., 1989; Andreasen et al., 1996). Borders between cytoarchitectonically defined thalamic nuclei can only be definitively determined post-mortem.
However, each thalamic nucleus has a distinct pattern of cortical and subcortical connectivity (Van Buren and Burke, 1972; Jones, 1985; Goldman-Rakic and Porrino, 1985; Yeterian and Pandya, 1985, 1988, 1991, 1997; Darian-Smith et al., 1990; Guillery and Sherman, 2002). We have previously exploited this fact to parcellate the human thalamus in vivo on the basis of its cortical connectivity (Behrens et al., 2003a). Our approach uses diffusion-weighted magnetic resonance images to perform probabilistic tractography (Behrens et al., 2003b). Diffusion-weighted imaging (DWI) is able to characterize the apparent diffusion properties of water in different directions (Basser et al., 1994). In an anisotropically organized structure such as the brain, local water diffusion is fastest along the predominant orientation of fibre tracts. DWI can therefore be used to infer these fibre orientations in white (Henkelman et al., 1994; Conturo et al., 1999; Jones et al., 1999b; Mori et al., 1999) or grey matter (Wiegell et al., 2003). By classifying thalamic voxels according to the cortical region with which they show the highest connection probability, we can parcellate the thalamus into distinct ‘connectivity-defined regions’ (CDRs) that we hypothesize correspond to thalamic nuclei or nuclear groups (Behrens et al., 2003a).
Here, we test the anatomical and functional validity of this correspondence more directly and begin to characterize the variability of CDRs across a healthy population. First, we produce group probability maps that express the population likelihood of connection to each cortical region for every point in the thalamus. Second, to test the correspondence between our CDRs and thalamic nuclei, we compare the relative volumes of our CDRs to cytoarchitectonically defined nuclear volumes from a previous histological atlas of the thalamus. Finally, we test the correspondence between our structurally defined thalamic parcellation and the localization of previously reported functional activations during motor or executive tasks. This not only serves to validate our approach, but also illustrates how information from the group probability atlas can be used to guide structure–function correlations.
Material and Methods
Diffusion-weighted (DW) data were acquired in 11 healthy subjects (7 male, 4 female, mean age 34.9 ± 11.2 years) using echo planar imaging with optimized cardiac gating (Wheeler-Kingshott et al., 2002) (60 × 2.3 mm thick slices, field of view = 220 × 220 mm2, matrix = 96 × 96; reconstructed on 128 × 128 matrix giving final resolution of 1.7 × 1.7 × 2.3 mm3) implemented on a General Electric 1.5 T Signa Horizon scanner with standard quadrature head-coil and maximum gradient strength of 22 mT/m. Informed written consent was obtained from all subjects in accordance with approval from the National Hospital for Neurology and Neurosurgery and the Institute of Neurology Joint Research Ethics Committee. The diffusion weighting was isotropically distributed along 54 directions (δ = 34 ms, Δ = 40 ms, b-value = 1150 s/mm; Jones et al., 1999a). Six diffusion-weighted volumes (b-value = 300 s/mm) and six volumes with no diffusion weighting were acquired.
In each subject, a high-resolution T1-weighted scan was obtained with a 3-D inversion recovery prepared spoiled gradient echo (IR-SPGR; FOV = 310 × 155; matrix = 256 × 128; in-plane resolution = 1.2 × 1.2 mm2; 156 × 1.2 mm thick slices; TI = 450 ms; TR = 2 s; TE = 53 ms).
Probabilistic diffusion tractography was carried out according to previously detailed methods (Behrens et al., 2003a,b). Using Bayesian techniques, we estimate a probability distribution function (pdf) on the principal fibre direction at each voxel. We then generate probability distributions of connectivity between seed and all other points by repeatedly sampling connected pathways through this pdf field. The effect of this procedure is to build a probability distribution on the location of the dominant connection from the seed voxel.
We manually outlined the whole thalamus (creating a ‘mask’) and seven exclusive cortical regions [prefrontal (PFC), primary motor (M1), premotor (lateral and medial) (PMC), temporal, posterior parietal (PPC), primary and secondary somatosensory (S1/S2) and occipital cortices; Fig. 1Aiii] on each subject's T1-weighted image using anatomical landmarks detailed previously (Behrens et al., 2003a). For tissue-type segmentation, skull stripping and registration, tools from the FMRIB Software Library (FSL; www.fmrib.ox.ac.uk/fsl) were used. We performed probabilistic tissue type segmentation and partial volume estimation on the T1-weighted image (Zhang et al., 2001) and masked cortical regions to include only voxels estimated at >15% grey matter. We skull-stripped (Smith, 2002) diffusion-weighted, T1-weighted and MNI standard brain template images (Evans et al., 2003) and performed affine registration (Jenkinson and Smith, 2001) to derive transformation matrices between the three spaces.
From each voxel in the thalamus mask, we drew samples from the connectivity distribution and recorded the number of samples that passed through each cortical mask. For each thalamic voxel in standard space we calculated a probability of connection to each cortical zone as a proportion of the total number of samples from that voxel that reached any cortical area. Hard segmentation of the thalamus was performed by classifying seed voxels as connecting to the cortical mask with the highest connection probability, resulting in exclusive connectivity-defined regions (CDRs). For each cortical area, we thresholded and binarized individual subject results to include only those thalamic voxels with a connection probability >25%. These images were overlaid to create group probability maps of thalamic sub-regions.
In order to derive relative volumes of cytoarchitectonically defined nuclei we manually measured the cross-sectional area of major nuclei delineated in axial cytoarchitectonic slices from a single human brain (Schaltenbrand and Wahren, 1977) and multiplied by slice thickness to calculate the volume of each thalamic nucleus. Many of our cortical target regions are connected to more than one thalamic nucleus and many nuclei connect to more than one cortical target. Therefore, in order to estimate cytoarchitectonic-based volumes for the total region of the thalamus projecting to each of the seven cortical target regions, we combined the nuclear volumes into groups as follows: M1 (ventrolateral plus centremedian nuclei), PFC (anterior, mediodorsal and ventral anterior nuclei), PMC (ventral anterior nucleus), PPC (lateral posterior plus pulvinar nuclei), S1/S2 (ventral posterior lateral nucleus), temporal (pulvinar nucleus). Half of the volume of the ventral anterior nucleus was assigned to PFC and half to PMC. Similarly, the volume of the pulvinar was equally split between PPC and temporal targets. The volume of the individual subject DWI-based CDRs for each cortical target was found directly from our imaging data and averaged across subjects giving a single mean volume for each CDR. As we were interested in relative rather than absolute volumes, both cytoarchitectonic- and DWI-based measures were normalized to their respective maximum volumes. To test directly the correspondence between the seven volumes (one for each cortical target area) derived from the cytoarchitectonic atlas and the seven defined from DWI, we calculated the correlation between them using Pearson's correlation co-efficient with a one-tailed probability threshold of P < 0.05.
We compared the locations of group probability maps to centres of thalamic activations from functional imaging studies of upper limb movements and of executive tasks that activated the motor/premotor and prefrontal cortices, respectively. Studies for inclusion were identified by searching the PubMed database (http://www.ncbi.nlm.nih.gov/entrez/). We included BOLD fMRI or rCBF PET studies published between 1 January 1999 and 30 December 2003 in which coordinates of whole brain activation were reported for a group of at least five healthy human control subjects. We included only studies published in the following journals: Neuroimage, Journal of Neuroscience, Cerebral Cortex, Proceedings of the National Academy of Science USA, Brain Research, Cognitive Brain Research, Neuron, Neuropsychologia, Journal of Cognitive Neuroscience, Brain and Experimental Brain Research. To identify motor tasks that activated motor or premotor cortex, we used the search term (fmri OR pet) AND ((motor OR sensorimotor OR premotor) AND cortex) AND movement AND (hand OR finger). Papers that reported activation for unilateral hand or finger motor tasks versus rest or a non-motor baseline, or activations that varied with a simple movement parameter (rate or force) were included. This identified a total of 38 studies that fulfilled our selection criteria. Of these, 24 did not report any thalamic activation for the contrasts of interest and 14 did report thalamic activations (of which two fell outside the thalamus). Twelve movement-related thalamic activations were therefore included in further analyses.
To identify executive/memory tasks that activated the prefrontal cortex we used the search term (fmri OR pet) AND prefrontal AND ((planning) OR (n back) OR (wisconsin card sorting) OR (prospective memory) OR (source memory) OR ((task OR set) AND (switch OR switching OR shift)) OR (delayed response)). Papers that reported activation for the cognitive tasks listed in this search time versus a control or rest baseline were included. This identified 59 studies that fulfilled our inclusion criteria. Of these, 41 did not report thalamic activation and 18 did report thalamic activation. The 18 studies with thalamic activation reported a total of 31 thalamic activation centres of which four fell outside the thalamus leaving 27 thalamic activation centres to enter into further analyses.
Coordinates from functional imaging studies are reported in MNI space (Evans et al., 2003) and in cases where the MNI template was not originally used coordinates were converted using a simple non-linear transform (http://www.mrc-cbu.cam.ac.uk/Imaging/mnispace.html).
Connectivity-based Anatomy: Probabilistic Segmentation
Classification of thalamic voxels based on the cortical region with which they showed the highest connection probability resulted in the definition of clusters of commonly-connected voxels whose relative locations correspond well to major thalamic nuclei (Fig. 1A). Qualitatively, the relative locations and extents of these individual connectivity-defined sub-regions (CDR) were reproducible across subjects (Fig. 1B). Although the relative positions of CDRs were preserved across subjects, there was some variability in the precise volumes and locations of borders (Fig. 1B). However, in these ‘hard’ segmentations, in which voxels are classified according to their highest cortical connection probability, the edges of clusters may depend on voxels that show only low probability of connection to any cortical region (e.g. from areas that have predominantly subcortical connections).
To characterize voxel-wise correspondence in thalamic connections across subjects quantitatively, we co-registered binarized masks of thalamic regions showing >25% probability of connection to each of the defined cortical areas. The resulting group probability maps are centred on localized regions of high probability (across the group) of connection to each cortical region (Fig. 1C). These results have been entered into a probabilistic atlas which is available at http://www.fmrib.ox.ac.uk/connect. Because the borders of these probability maps are intrinsically fuzzy, we used thresholded group maps to conveniently display boundaries between the sub-regions of the thalamus in a standard brain space (Fig. 2).
Having established evidence for between-subject reproducibility of our thalamic parcellation, we further tested the hypothesis that the sub-regions found in this way (i.e. the individual subject CDRs) correspond to thalamic nuclei or nuclear groups. The correspondence between the relative locations of our CDRs and cytoarchitectonically defined nuclei in the thalamus post-mortem (Fig. 1A) supports this claim. We also compared the mean relative volumes of individual CDRs and cytoarchitectonically defined nuclear groups. We found a strong correlation between the mean relative volume across subjects defined using DWI and those based on a cytoarchitectonic atlas (Fig. 3A, left hemisphere r = 0.71, P = 0.038; right hemisphere r = 0.70, P = 0.04).
Overlapping Connections to Multiple Cortical Areas
The analysis thus far has considered only the primary cortical connection of each CDR. However, many of the voxels within CDRs generated connections to cortical regions other than their primary cortical target. Quantitative expression of connection probability allows us also to characterize anatomically overlapping connectivities. The connectivity profiles of CDRs are highly reproducible across hemispheres and demonstrate that when CDRs show connectivity to multiple cortical regions these regions tend to be functionally related and/or physically adjacent (see online Supplementary Material). For example, the CDR with primary motor connections also has strong connections to premotor and somatosensory cortices. The existence of multiple cortical connections from each CDR could reflect multiple factors. A likely contributing factor is the fuzzy borders between cortical regions. However, an additional factor might be genuine interdigitation of regions within the thalamus connecting to disparate cortical areas. This could, for example, reflect the different distributions within the thalamus of distinct cell types with diffuse and specific cortical connections — so-called matrix and core neurons (Jones, 2001). We found evidence for interdigitation in the distribution of thalamic connections to PFC. Examination of the group mean probability map for connections to M1 and PFC from the thalamus demonstrates that the region connecting to PFC does not simply blur into the region connecting to M1. Rather, there are two distinct areas with PFC connections — a primary one that probably includes the mediodorsal nucleus and a secondary region, located within a sub-region anatomically corresponding to the ventrolateral nucleus (Fig. 3B). To assess the reliability of this pattern of PFC connection apparent in the mean connection probability profile, we inspected the corresponding slice in all individual subjects and found that a pattern of dense medial PFC connection, followed by a region without PFC connection, followed by a second, more lateral/inferior region of PFC connection was present in six out of eleven subjects.
Co-localization with Centres of Functional Activation
We have proposed that the atlas based on our group probability maps could be used to assign likely anatomical labels to sites of brain activation or lesions. To test directly the functional–anatomical validity of the boundaries defined by our group maps, we assessed the correspondence between connectivity-defined volumes and previously reported centres of gravity of functional activations localized to the thalamus. Functional activations reported during motor and executive/memory tasks tended to cluster in distinct regions of the thalamus (Fig. 3Ci,ii). We used the atlas based on our group probability maps to characterize the connection probability profile for each functional activation centre (Fig. 3-D). Activation centres for the motor tasks tended to fall within regions of high probability of connection to sensorimotor and premotor areas (Fig. 3Ciii,iv,D). Similarly, activation centres for the executive and memory tasks fell within thalamic regions of high probability of connection to prefrontal cortex (Fig. 3Cv,vi,D).
We provided evidence in support of the functional and anatomical validity of connectivity-based segmentations of the thalamus and characterized their variability. Previously we showed that diffusion tractography can be used to parcellate the thalamus into regions that we propose correspond to nuclei or nuclear groups (Behrens et al., 2003a). Parcellations of the human thalamus based on cortical connectivity fit well with predictions based on invasive tract tracing studies on thalamo-cortical and cortico-thalamic connectivity in non-human animals (Jones, 1985; Goldman-Rakic and Porrino, 1985; Yeterian and Pandya, 1985, 1989, 1991, 1997; Darian-Smith et al., 1990, 1996). Here we have defined the variation of human thalamic parcellation between individuals quantitatively by generating a probabilistic thalamic atlas. This was used to test more rigorously the hypothesis that the parcellation corresponds to functional anatomical divisions within the thalamus. First, we found good agreement between volumes of the thalamic sub-regions obtained using our method and comparable sub-regions from previous cytoarchitectonic data. Second, the locations of the individual regions correspond well with data from prior functional imaging experiments. We showed, for example, that thalamic activations with motor paradigms map into a region corresponding to the ventral lateral nucleus, while activations associated with tasks involving executive control and working memory co-localized with the connectivity-defined sub-region including the mediodorsal nucleus.
Parcellation of brain regions with respect to connectivity offers a complementary tool to conventional imaging of anatomical boundaries using tissue contrast arising from, e.g. differences in myeloarchitecture (Magnotta et al., 2000). A notable feature of the current thalamic parcellation is that borders between regions are not sharp. In individual subject parcellations this may be due in part to genuinely fuzzy borders between thalamic regions or their cortical targets (e.g. connectivity to primary and premotor cortices from the ventral lateral and ventral lateral anterior nuclei). In addition, for many of the major relay nuclei of the thalamus, inputs are predominantly subcortical rather than cortical and even for nuclei with strong cortical connectivities, the distribution of these connectivities is based on cortical units (e.g. cytoarchitectonic regions, cortical layers) smaller than the large cortical regions used here (Goldman-Rakic and Porrino, 1985; Jones, 2001). Therefore, nuclear boundaries cannot always be expected to precisely match boundaries defined solely on the basis of cortical connectivity at the level used here.
Methodological factors, such as the limited spatial resolution and motion artefacts caused by pulsation of the cerebrospinal fluid in the third ventricle, will also contribute to the fuzziness of borders in our parcellation. Note, however, that as the tractography algorithm considers diffusion measurements not only within a voxel but also at a more global scale (Behrens et al., 2003b), the effective resolution of our parcellation will be finer than that of the original diffusion images. The lack of sharp borders in the group probability maps will additionally reflect inter-individual variability in thalamic anatomy. Previous cytoarchitectonic studies have demonstrated variability in nuclear size and location between individuals (Morel et al., 1997). Inter-individual variation in connectivity-defined parcellations will also reflect difficulties in precisely matching variations in brain and thalamic sizes and shapes in registration of images across the group. Finally, even within an individual, distortions and low spatial resolution in the EPI images limit the degree to which the high resolution structural and diffusion tractography images can be matched spatially. Here, we used a linear registration algorithm (Jenkinson and Smith, 2001). It could be argued that linear registration is preferable if the aim is to characterize anatomical variability. However, an alternative approach would be to use non-linear methods to match the external borders of the thalamus across subjects and then focus on variability in connectivity-defined intrinsic boundaries.
It has previously been difficult to validate directly diffusion tractography due to a lack of alternative methods providing similar data in humans, as well as the relative paucity of animal imaging data. Here we took two approaches to validation of the connectivity-based parcellation: first by comparison to cytoarchitectonic atlases and second by comparison to functional activations.
The demonstration that relative locations and volumes defined using a connectivity-based parcellation of the thalamus generally correspond well with volumetric measurements from cytoarchitectonic maps is a powerful test of the ability of the method to define correct targets and, by implication, to identify appropriate paths. We would not expect the absolute volumes to correspond between these two methods. While cytoarchitectonic measurements were made up only of volumes from major nuclei, the measurements based on diffusion imaging would have included contributions from smaller nuclei and from white matter regions within the thalamic volume. Therefore, not unexpectedly, the absolute volumes of sub-regions are greater for the connectivity based measures. Differences in relative volumes may also reflect technical factors that will result in varying sensitivity to different anatomical pathways. The diffusion tractography approach used here is sensitive primarily to major pathways and therefore smaller pathways, pathways with sharp path inflections or pathways that cross other tracts are not always detected. This may explain the relatively small size of the motor nucleus by our method compared to cytoarchitectonic measurements: pathways reaching the lateral parts of the motor strip from the thalamus would have to cross the superior longitudinal fasiculus and therefore may not be detected. Future work should assess quantitatively the correspondence between the locations of connectivity-defined CDRs and thalamic nuclei by comparison of centres and extents of such regions in standard stereotactic space. Such comparisons, however, are not straightforward, as there is unlikely to be a one-to-one correspondence between individual nuclei and CDRs (e.g. the CDR connecting to the posterior parietal cortex is likely to include regions of both the lateral posterior nuclei and parts of the pulvinar.)
Relating the thalamic parcellation to locations of centres of functional activations offers an alternative approach to its validation (and also illustrates an important potential application of the parcellation method). Although thalamic activation is frequently reported in imaging studies (Baker et al., 1996; Davis et al., 1998; LaBar et al., 1999; Dove et al., 2000; McGonigle et al., 2000; Gerardin et al., 2003), authors rarely assign activations to a specific nucleus, due to the problems of inferring nucleic architecture with current approaches. Here we have chosen to test the functional validity of connectivity-defined volumes with data from well-characterized tasks to minimize confounds in interpretation. Centres of activation associated with two distinct types of tasks showed good functional–anatomical correspondences: motor activations co-localized to the sub-region corresponding to the ventrolateral nucleus and activations during executive tasks were mainly located within the sub-region corresponding to the mediodorsal nucleus. The atlas based on our group maps presented here can be used to assign a probabilistic anatomical label to activation foci within the thalamus more generally. This could prove particularly useful in cases where functional connectivity between the thalamus and neocortex is otherwise ambiguous. For example, although thalamic activation is frequently reported in studies of pain (Davis et al., 1998; Becerra et al., 1999; DaSilva et al., 2002), the nature of the thalamic processing remains unclear and would be illuminated substantially by a clarification of precisely which thalamic nuclei are involved in processing nociceptive versus non-nociceptive stimulation in the human brain.
The ability to perform probabilistic grey matter parcellation in vivo also raises new possibilities for the study of the brain in disease. The fact that connectivity-based segmentation of the thalamus is reproducible across a group of healthy individuals demonstrates the feasibility of this approach for study of disorders with putative thalamic pathology. Connectivity between the thalamus and prefrontal cortex is hypothesized to play a role in schizophrenia, for example (Andreasen et al., 1996). The relative size and location of the thalamic sub-region with a high probability of connection to prefrontal cortex could be directly compared between schizophrenic subjects and healthy controls. Lesions of the thalamus itself can cause a variety of cognitive impairments and it has been suggested that particular impairments relate to specific nuclear damage, e.g. lesions to the mediodorsal nucleus are associated with executive dysfunction whereas lesions involving the intralaminar nuclei lead to impairments in attention (Van der Werf et al., 2003). Localizing lesion sites on the group probability maps presented here would enable the generation of hypotheses concerning the likely thalamo-cortical pathways affected and enable more precise clinico-anatomical correlations to be made.
Supplementary material can be found at: http://www.cercor.oupjournals.org/
We acknowledge the support of the Wellcome Trust (H.J.B.), UK Medical Research Council (P.M.M.), UK Engineering and Physical Science Research Council (T.E.J.B., S.M.S.) and Multiple Sclerosis Society of Great Britain and Northern Ireland (P.M.M.). We are grateful to D. Mortimer and D. Flitney for assistance with figures and the web-based atlas, K. Sheehan for masking, C. Wheeler-Kingshott, G. Barker and P. Boulby for data acquisition, J.M. Brady for invaluable supervision (T.E.B.) and an anonymous reviewer for useful suggestions for improving the manuscript.
1Centre for Functional Magnetic Resonance Imaging of the Brain, University of Oxford, John Radcliffe Hospital, Oxford OX3 9DU, UK and 2Institute of Neurology, Queen Square, University College London, London WC1N 3BG, UK