It is generally agreed that the cerebral cortex can be segregated into structurally and functionally distinct areas. Anatomical subdivision of Broca's area has been achieved using different microanatomical criteria, such as cytoarchitecture and distribution of neuroreceptors. However, brain function also strongly depends upon anatomical connectivity, which therefore forms a sensible criterion for the functio-anatomical segregation of cortical areas. Diffusion-weighted magnetic resonance (MR) imaging offers the opportunity to apply this criterion in the individual living subject. Probabilistic tractographic methods provide excellent means to extract the connectivity signatures from diffusion-weighting MR data sets. The correlations among these signatures may then be used by an automatic clustering method to identify cortical regions with mutually distinct and internally coherent connectivity. We made use of this principle to parcellate Broca's area. As it turned out, 3 subregions are discernible that were identified as putative Brodmann area (BA) 44, BA45, and the deep frontal operculum. These results are discussed in the light of previous evidence from other methods in both human and nonhuman primates. We conclude that plausible results can be achieved by the proposed technique, which cannot be obtained by any other method in vivo. For the first time, there is a possibility to investigate the anatomical subdivision of Broca's area noninvasively in the individual living human subject.
It is widely accepted among neuroscientists that the cerebral cortex can be subdivided into a number of structurally and functionally distinct areas. This subdivision is referred to by the term “cortex parcellation” throughout this paper. Under the premise that structure reflects function, understanding the structural organization of the cortex is vital for the study of brain function. However, what are the parameters that define a cortical area as structurally and functionally uniform and distinct from its neighbors?
Structural interpretations of functional imaging data are mostly based on the cytoarchitectonic map of Brodmann (1909), which reflects the specific variation in size and packing density of cell bodies over the layers of the cortical sheet in one single subject. Other parameters have been added successively, such as distribution and amount of intracortical myelinated fibers (Vogt 1910, 1911; Braitenberg 1962) or the density of certain neurotransmitters receptors (Zilles and others 2004). All these parameters describe the local microarchitecture of cortical tissue. However, as has been pointed out by Kaas (2002) and others (cf., Lashley and Clark 1946), differences among the microarchitectures of cortical areas are often subtle. In fact, it has been demonstrated that previously cyto- and myeloarchitectonically indistinguishable areas show a clear functional segregation (cf., Roland and Zilles 1998; e.g., Friederici and Kotz 2003, for the superior–posterior and the inferior portion of Brodmann area [BA] 44). Some of these functional segregations could be confirmed by receptorarchitectonic maps, for example, the separation of BA44 into a dorsal and a ventral portion supporting phonological and syntactic processes, respectively (Friederici 2002; Amunts and Zilles 2006). Furthermore, it has been argued that differences in input from and output to other brain areas may be a necessary additional criterion for the segregation of functionally different areas. (In general, a subdivision based on more than one type of evidence would be more reliable; cf., van Essen 1985; Kaas 1997.) Hence, the pattern of anatomical connectivity (within this article, we are limited to long-range connectivity, i.e., corticocortical association fibers, the corticospinal tract, transcallosal connections, etc. Intra-area connections or fibers among adjacent areas [u-fibers] are beyond the scope of this paper) seems to be an important parameter for the description and distinction of cortical areas (cf., also Barbas and Rempel-Clower 1997). This is particularly plausible in the understanding that higher cognitive functions are preferentially based on widespread networks, rather than isolated cortical areas. Therefore, the anatomical characterization of a particular cortical area just by its internal microstructure without regarding its connections to other brain areas must remain incomplete.
To date, connectivity information has been revealed mostly from animal models. They focus on the measurement of degenerating axons subsequent to lesion or using active transport of tracers that are injected while the animal is still alive (Passingham and others 2002). Though powerful, active axonal transport techniques have seen little or no application in human brain because they are limited to invasive studies of living tissue. In principle, “postmortem” tracer application allows tracing of tracts too, in aldehyde-fixed (Haber 1988) as well as in unfixed (McConnell and others 1989) tissue but only for distances of about 10 mm (Mufson and others 1990). Longer distance connections can only be investigated by dissection studies (Klingler and Gloor 1960), by indirect evidence from anterograde degeneration due to brain lesions (DiVirgilio and Clarke 1997), or with optical imaging approaches based on series of histological sections throughout the respective tract (e.g., Axer and others 2002). Due to these restrictions, information about long-range connectivity in the normal human has been difficult to obtain and is limited in scope (Clarke and Miklossy 1990; Miklossy and others 1991; Crick and Jones 1993). With the advent of modern imaging technology, it has become possible to quantify long-range connectivity in vivo on the basis of diffusion-weighted magnetic resonance imaging (DW-MRI) using white matter tractography.
DW-MRI measures the direction-dependent mobility of water molecules (diffusion), which is influenced by the microscopic architecture of the brain tissue (Pierpaoli and others 1996; Le Bihan and others 2001; Beaulieu 2002), yielding information on the local orientation of the white matter fibers. White matter tractography attempts to integrate the local fiber orientation information by tracing the course of white matter fiber bundles throughout the brain (for a review, see Mori and van Zijl 2002). Two major approaches are used to tackle this task: deterministic and probabilistic tractography. Deterministic tractography produces maximum likelihood pathways through the diffusion-weighted magnetic resonance (MR) data set. Probabilistic approaches to the tractography problem take into account the uncertainty of the direction information in each voxel (Koch and others 2002; Behrens, Woolrich, and others 2003). Commencing at some seed voxel, tracking yields a probabilistic distribution of possibly connected voxels—the so-called “tractogram.” The basis for probabilistic tractography is the local probability density function (lPDF) generated by a mathematical model of the local diffusivity, for example, the diffusion tensor model (Basser and others 1994). Here, each of the model parameters is characterized by a statistical distribution (e.g., expectation value and standard deviation of a normal distribution). In a second step, the lPDFs have to be connected using conditional probabilities.
The most straightforward application of tractography is the segmentation of the telencephalic white matter into compartments corresponding to the various systems of fiber bundles (Catani and others 2002). The other obvious application concerns the degree of anatomical connectivity among different gray matter areas (Koch and others 2002; Behrens, Woolrich, and others 2003). Finally, if one seeks to subdivide brain structures according to their connectivity patterns, one can rely on the discriminative power of tractographic techniques. This has led to another powerful application of white matter tractography, namely connectivity-based cortex parcellation. The concept for a cortex parcellation based on tractography has been put forward by Johansen-Berg and others (2004) and has been demonstrated to yield promising results for the medial frontal cortex (supplementary motor area [SMA]/pre-SMA, Johansen-Berg and others 2004) and the geniculate bodies (medial vs. lateral, Devlin and others 2005).
In our study, we demonstrate that connectivity-based cortex parcellation is a suitable means to subdivide Broca's area, an area of high anatomical and functional complexity. For this purpose, we developed a technique, which combines a 3-dimensional (3D) extension of the tractographic method of Koch and others (2002) with some methodological concepts proposed by Johansen-Berg and others (2004). To demonstrate that this method can be reproduced, we first replicated previous findings (parcellation of the medial frontal cortex). Then, we applied the method to Broca's area.
Broca's area resides in the inferior frontal cortex (IFC) and traditionally comprises the lateral triangular and opercular part (“pars triangularis” and “pars opercularis”) of the inferior frontal gyrus (IFG). Here, not only these surface portions were included into the definition of Broca's area but also the base of its opercular part, that is, cortical band between the crown of the opercular part of IFG and the anterior insula. We refer to this structure as “deep frontal operculum” (Fig. 1).
It is well known that Broca's area can be structurally segregated using cytoarchitectonic methods (von Economo 1929; Sanides 1962, 1966; Amunts and others 1999; Amunts and Willmes 2005; Amunts and Zilles 2006) (dysgranular BA44 [roughly corresponding to opercular IFG], granular BA45 [roughly corresponding to triangular IFG]) or receptor mapping (Zilles and others 2002; Amunts and Zilles 2006). However, in vivo separation in individual subjects has not yet been achieved. We hypothesize that the subregions of Broca's area are distinct not only in their microanatomy but also in their pattern of long-range connectivity. If valid, connectivity-based parcellation would offer the possibility to define their borders in the “individual living subject” on the basis of the connectional architecture of neuronal networks, rather than macroanatomical features, like gyri and sulci. For a more general reflection on relating connectional architecture to gray matter function by using diffusion tensor imaging (DTI), see Behrens and Johansen-Berg (2005).
Data Acquisition and Preprocessing
Diffusion-weighted data and high-resolution 3-dimensional (3D) T1-weighted as well as 2-dimensional T2-weighted images were acquired in 6 healthy right-handed subjects (28 ± 3 years, 3 females) on a Siemens 3-T Trio Scanner with an 8-channel array head coil and maximum gradient strength of 40 mT/m. The diffusion-weighted data were acquired using spin-echo echo planar imaging (EPI) (time repetition [TR] = 8100 ms, echo time [TE] = 120 ms, 44 axial slices, resolution 1.7 × 1.7 × 3.0 mm, 10% gap, 2 acquisitions). Written informed consent was obtained from all subjects in accordance with the ethical approval from the University of Leipzig. The diffusion weighting was isotropically distributed along 24 directions (b value = 1000 s/mm2). Additionally, a data set with no diffusion weighting was acquired. The high angular resolution of the diffusion weighting directions improves the robustness of probability density estimation (see below) by increasing the signal-to-noise ratio per unit time and reducing directional bias. The total scan time for the diffusion-weighted imaging (DWI) protocol was approximately 7 min.
As first step in preprocessing the data, the 3D T1-weighted (magnetization prepared–rapid gradient echo; TR = 100 ms, time to inversion = 500 ms, TE = 2.96 ms, resolution 1.0 × 1.0 × 1.0 mm, flip angle 10°, 2 acquisitions) images were reoriented to the plane of anterior and posterior commissure. Upon reorientation, the T2-weighted images (pulse sequence: rapid acquisition with relaxation enhancement, RARE; TR = 7800 ms, TE = 105 ms, 44 axial slices, resolution 0.7 × 0.7 × 3.0 mm, 10% gap, flip angle 150°) were coregistered to the reoriented 3D T1-weighted images according to a “mutual information” registration scheme (Studholme and others 1997). Subsequently, by using in-house software (Wollny and Kruggel 2002), the DWI–EPI images were nonlinearly warped onto the T2-weighted images to reduce distortion artifacts. Following correction of image distortion, the diffusion tensor was calculated for each voxel using multivariate linear regression after logarithmic transformation of the signal intensities (Basser and others 1994). The fractional anisotropy (FA) of the tensor in each voxel was subsequently determined, and a multislice FA image (Basser and Pierpaoli 1996) was created.
For presentation purposes, on basis of the T1-weighted images, cortical surfaces have been rendered by using Freesurfer (Dale and others 1999).
Definition of the Regions of Interest
Regions of interest (ROIs) for the white matter tractography in each subject were defined by an expert (DYvC) on the basis of individual T1-weighted magnetic resonance (MR) images. They encompassed parts of the left IFC, that is, the deep frontal operculum as well as the surface portion of the opercular and triangular part of the IFG (Fig. 1). The ROIs were segmented into white and gray matter compartments thresholding the FA (white matter: FA > 0.1). Each white matter voxel in the ROI that neighbored a gray matter voxel was labeled as a seed voxel for the tractographic procedure (Fig. 2).
White Matter Tractography
We developed a 3D extension of the random walk method proposed by Koch and others (2002). The algorithm was applied to each of the seed voxels. The target space was the whole white matter volume with a resolution of 1.0 × 1.0 × 1.0 mm3. The algorithm can be described by a model of randomly moving particles. Imagine a particle in a seed voxel A, moving in a random manner from voxel to voxel. The transition probability to a neighboring voxel depends on the lPDF based on the local diffusivity profile that is modeled from the DTI measurement. This lPDF is discretized into 26 directions corresponding to neighboring voxels and yields higher transitional probabilities along directions with high diffusivity, that is, the presumed fiber directions. Hence, the particle will move with a higher probability along a fiber direction than perpendicular to it. If we perform this “experiment” many times and count how often particles from voxel A reach a target voxel B, we obtain a relative measure of the probability of tracing a pathway between the 2 voxels. The random walk is stopped when the particle leaves the white matter volume. The 3D distribution of the connectivity values of a particular seed voxel with all voxels in the brain is called a tractogram. The average tractogram for an entire region we call the “tractographic signature” of this region.
The exponent a = 7 is used to focus the probability distribution to main fiber direction and suppress the influence of the transverse diffusion. The value was empirically chosen in such a way that the trajectories of most particles follow the main fiber directions as defined by the lPDF. The transition directions in the local model are limited to the 26 discrete neighbors of the voxel, which is sufficient to produce a smooth distribution of the fibers directions after interpolation of the tensor data to 1 mm voxel size. A total of 100 000 particles were tested for each seed voxel. To compensate for the distance-dependent bias, each connectivity value is normalized to the shortest pathway (the “shortest pathway” between seed and target voxel refers to the smallest number of jumps that any one out of the 100 000 test particles has needed to reach the particular target) between the seed and the respective target voxel. After reducing the dynamic range of the connectivity values by logarithmic transformation, the entire tractogram was scaled to its maximum value. To remove random artifacts, only connectivity values bigger than 0.4 were used for further processing.
The idea behind cortex parcellation is that cortex areas with similar long-range connectivity are combined into a region, which is segregated from neighboring regions with different connectivity. The connectivity pattern of a cortical voxel is approximated by the tractogram associated to its neighboring white matter voxel. The general principle of the parcellation technique is illustrated in Figure 2. For each seed voxel within a ROI, a tractogram was computed. The correlation values between any 2 of these tractograms were computed and arranged in the “connectivity correlation matrix.” The ith column or row in this symmetric matrix characterizes the degree of similarity between the tractogram of seed voxel i and all other seed voxels. Consequently, the very high dimensional space of tractograms (dimension: number of voxels in the white matter) is transformed to a lower dimensional space of tractogram similarity patterns (dimension: number of seed voxels in the ROI). It is immediately plausible that voxels with similar connectivity patterns (i.e., tractograms) show also a similar pattern of correlation with the tractograms belonging to other seed voxels (i.e., similar columns in the connectivity correlation matrix). In order to define collections of voxels with similar connectivity, a multidimensional k-means cluster algorithm was applied to the columns of the connectivity correlation matrix. Like for any clustering method, one pitfall is the trade off between model consistency (how well does the clustering describe the structure of the data) and model complexity (preference of a simple model that describes the relevant features and ignores noise), with the consequence that the number of clusters has to be introduced from outside. Therefore, different numbers of clusters can be tested. We accepted only those solutions that were consistent across subjects, that is, the principal arrangement of the areas associated with the clusters was the same. Moreover, each area had to represent a single coherent region of the cortex. Finally, every area was additionally characterized by its tractographic signature (tractographic signature: average tractogram for a certain region).
To demonstrate that connectivity-based cortex parcellation can be reproduced, the method was applied to the same region as investigated in the work of Johansen-Berg and others (2004)—the medial frontal cortex. A separation along the anterior commisure (AC) line into an anterior and a posterior part was found (Fig. 3A), which corresponds exactly with the findings of Johansen-Berg and others (2004). Increasing the number of clusters, the region was divided posterior–anteriorly into 3 areas. They presumably represent pre-SMA, SMA, and M1. Both the most anterior (pre-SMA) and the most posterior (M1) regions could be separated between the hemispheres, whereas the middle one (SMA) could not be split, even if the number of clusters was increased further (Fig. 3B). Most importantly, the separation between pre-SMA and SMA remained at the AC line. The tractographic signatures for the different areas largely resemble the results reported by Johansen-Berg and others (2004), who found that the posterior areas predominantly exhibit connections to premotor/motor regions and the corticospinal tract, whereas the anterior area seems to be more connected to prefrontal and parietal regions (Fig. 3C). We conclude that the replication of the results by Johansen-Berg and others (2004) with a different method documents the reproducibility of this kind of approach to cortex parcellation.
In all 6 cases, the method revealed a clear segregation of Broca's area into the surface portion of BA44/45 and the deep frontal operculum (Fig. 4). Additionally, in 5 subjects, the surface portion split along the ascending branch of the lateral fissure into a rostral and a caudal part, suggesting a separation between BA44 and BA45. In the remaining data set (subject III), the clustering algorithm left the surface portion of the IFG undivided. In subject I, the ventralmost portion of the triangular part forms a common region together with the deep frontal operculum. A similar effect is observed in subject III for the posterior convexity of the orbital part of the IFG.
The tractographic signatures of the deep frontal operculum suggest projections through the anterior subinsular white matter and the inferior level of the external/extreme capsule (note that the DTI data do not provide sufficient resolution to distinguish between the external and the extreme capsule, which are separated by the claustrum) to the temporal isthmus. From there, the signatures can be further traced to the anterior temporal lobe, into putative inferior occipitofrontal and inferior longitudinal fascicle. Moreover, a second more inferior (ventral) pathway seems to follow the uncinate fascicle in lateral direction into the anterior temporal lobe in all but one subject. Medially, this pathway can be traced into the basal forebrain, that is, into the amygdala complex and into the septal region. In 2 subjects, there are additional tracts to the frontopolar cortex.
Signatures associated with putative BA44 and BA45, the same pathway through the inferior external/extreme capsule as found for the deep frontal operculum. Additionally, a more superior (dorsal) pathway can be traced to the parietal lobe and toward the perisylvian region putatively via the arcuate and the superior longitudinal fascicle (in 4 subjects) as well as in all subjects to the dorsomedial prefrontal cortex and to the ventral portion of precentral gyrus. In the signatures of BA44, this dorsal pathway is more dominant, whereas for BA45, tracts into the dorsomedial prefrontal cortex and through the external/extreme capsule are more prominent. For the signatures of a typical subject (subject I), refer to Figure 5.
As reasoned in the introduction and evidenced by our findings, long-range connectivity is a functio-anatomically relevant trait of the brain. It is therefore a suitable criterion for the identification of mutually distinct cortical areas. Yet, it is not clear to what extent diffusion-based white matter tractography reflects the fine details of anatomical connectivity. (This also implies the question for the comparability with connectivity data obtained by alternative techniques, e.g., tracer studies.)
Resolution in the ROI
The proposed parcellation method relies on the differentiation of tractograms associated with all voxels included in the ROI. However, the limited spatial resolution of the measurement essentially restricts the differentiability of adjacent tractograms. Figure 6 illustrates the relationship between anatomical structures and diffusion tensor-magnetic resonance imaging (DT-MRI) voxel resolution. DT-MRI can image these structures only incompletely: 1) structure is not observed directly but only indirectly through water diffusion; 2) the dimensions of fine microanatomical structures (single cells, fibers, etc, which are in the order of microns) are far beyond the resolution currently available to MR-based techniques (i.e., in the order of millimetres) and cannot be resolved by tractography. Any tractographic result is based on the integrative effect of a large number of microscopic structures, including neural and glial tissue. Cortical voxels integrate information of structures oriented in many different directions and therefore do not exhibit much net anisotropy. (This is true for current DTI resolution, which cannot resolve e.g., the Baillarger stripes.) Directly underneath the cortical layers, there is the layer of u-fibers (about 0.75–1 mm thick, cf., Rockland 1989), which is connecting neighboring cortical areas. Only beyond this u-fiber belt, the actual long-range fiber bundles are sufficiently dominant to determine the direction of the gross diffusion within a voxel. Therefore, the seed points for our DTI tractography algorithm reside at the gray matter/white matter interface to ensure that the computed tractographic signatures are dominated by long-range connections, within the given resolution. In contrast, Johansen-Berg and others (2004) used both cortical and subcortical seed voxels. However, starting the tracking procedure in gray matter will allow the pathways to make detours over the neighboring cortical voxels, the transition to which is (almost) equally probable as to white matter fiber bundles. This might result in an additional loss of resolution.
However, even when seeding in the subcortical white matter, resolution is limited. DTI voxels sizes are currently in the order of 2–3 mm (Fig. 6). Hence, relatively few potential tractography seed voxels reside within a gyrus. Consequently, they bear just an approximate relationship to the cortex. Additionally, they may be subject to partial volume effects from u-fibers, resulting in similar tractograms for neighboring areas connected by u-fibers.
The precise relation between DTI-based tractography findings and the underlying microstructure remains unclear. The exact modeling of this relationship would be immensely complex and has not yet been accomplished. Moreover, even if such a model would exist, its inversion (i.e., the reconstruction of the microstructure from the DTI data) is an underdetermined problem with no unique solution. Hence, some of the identified connections might just be artifacts of the method. A major source of such tracking artifacts is created by close contact between parallel and crossing fiber bundles, causing the main fiber directions in the respective voxels to be poorly defined (Wiegell and others 2000; Barrick and Clark 2004). Strategies to combat this uncertainty have been proposed (for a review, see Mori and van Zijl 2002). However, there is no way to reconstruct information, which is not contained in the measured data. Therefore, instead of trying to resolve the irresolvable, one can quantify the uncertainty using probabilistic methods (Koch and others 2002; Behrens, Woolrich, and others 2003). This is the approach we followed in the present work. It avoids systematic biases to a certain extent and instead, introduces a quantification of uncertainty, where a clear decision on the propagation direction is impossible. We therefore did not obtain one single solution, as with deterministic tractography, but the so-called tractogram, which is a collection of possibly connected brain areas, weighted by their probability. An exact probabilistic (Bayesian) solution for tracking was proposed by Behrens, Johansen-Berg, and others (2003; Behrens, Woolrich, and others 2003). However, the accuracy of this global approach depends on the local model linking the fiber direction distribution to the measured diffusion profile. Because such local models are currently still quite imperfect, this global solution cannot unfold its full potential yet. Therefore, we opted for a more heuristic approach, similar to the one proposed by Koch and others (2002). In particular, the strong decay of the connectivity values with increasing distance from the seed voxel was counterbalanced by normalization to the path length. Therewith, a disproportional influence of the shorter range connectivity onto the correlations among the tractograms was avoided.
The employed algorithm has some parameters that are chosen heuristically. First, there is the exponent a (see eq. 1), which influences the translation between the local diffusion tensor and the lPDF. This parameter has clearly an influence onto the diversity of tracts found by the algorithm. The lower the exponent, the higher is the chance that the algorithm follows direction other than the main one, leading to more complete information, but also to a higher chance for artifacts. However, computations with different exponents show that the exact choice of a has a surprisingly small influence onto the parcellation result over a broad range of values. Another heuristic approach concerns the normalization to path length as well as the logarithmic transformation of the final connectivity values. This influences the relative importance of areas near and distant from the seed region onto the correlation among tractograms. The present choice makes sure that all regions (except for the immediate vicinity of the seed voxel) have approximately the same influence.
From above, one can conclude that the tractograms reflect anatomical connectivity only in an indirect way. However, for the purpose of cortex parcellation, not the connectivity itself, but its difference among areas is important. Hence, even not completely plausible tractograms can yield a correct parcellation result. It is certainly possible that also connectivity differences are eradicated by the described artifacts. Then, areas with different connectivity patterns are not distinguishable by their tractograms any more. On the other hand, established differences among tractograms—as we found them between the subregions within the IFG—allow the conclusion that there is also different connectivity.
Once the tractograms have been computed for every seed voxel in the ROI, it remains to be decided which of them are similar enough to be clustered together in one area. This is essentially a modeling process. We seek to describe a real situation, where in fact every seed voxel is associated with a different tractogram (i.e., with a simplified model). As with any model, there are the opposing demands for describing the data accurately (accounting for as much variance as possible) and effectively (with as few parameters as possible) at the same time. Once we have opted for a certain level of model complexity (i.e., the number of areas), we need to assign the seed voxels to these areas (i.e., determining the area boundaries) as objectively and reproducibly as possible. For this task, Johansen-Berg and others (2004) chose a method based on visual subdivision of the reordered correlation matrix of the tractograms. We replaced this approach by an automatic clustering algorithm. This gave us an objective way to assign seed voxels (i.e., tractograms) to the areas, once we have decided on the number of clusters.
Yet, the decision on the number of clusters is not easy to make. There might even be several correct choices, that is, several levels of parcellation. Hierarchical clustering methods could make a contribution here. The decision, which number(s) of areas describe the data better than others, could possibly be tackled by the use of information criteria, which attempt an information theoretically founded optimal compromise between model complexity and model accuracy (see e.g., Knösche and others 1998).
So far, structural subdivisions of IFC were only available from postmortem specimens. In living subjects, they had to be identified indirectly by macroanatomical landmarks. However, although there is often a good correlation of these landmarks with cytoarchitectonically defined borders of cortical areas, this relation is not necessarily a perfect one (Zilles and others 1997; Geyer and others 1999, 2000; Amunts and others 2000; Grefkes and others 2001), in particular not in the IFC (Amunts and others 1999, 2004). Furthermore, only the surface portions of the IFC are contained in cytoarchitectonic maps; its opercular base, the deep frontal operulum, is unfortunately not yet covered.
We have shown that tractography-based cortex parcellation can reliably segregate putative BA44 from the deep frontal operculum (Fig. 4). Moreover, in all but one case, also the connectivity-based separation between putative BA44 and BA45 was found to be aligned with the chosen macroanatomically identified boundary (i.e., the ascending branch of the lateral fissure). In the remaining case (subject III), only one area, covering both the opercular and triangular parts of the IFG, could be identified. However, the most anterior portion in this case was classified by its tractographic signature as putative BA45. This means that in this particular case, the boundary between BA44 and BA45 is located more anterior than macroanatomy would suggest. In another case (subject I), the most anterior portion of the triangular part is separated from putative BA45. It is characterized by a tractographic signature similar to the one attributed to the deep frontal operculum. We propose that in this case, the border between BA45 and BA47 is more posterior than macroanatomical landmarks (i.e., anterior branch of the lateral sulcus) would suggest. This proposal is paralleled by the macaque literature (Pandya and Yeterian 1996; Romanski, Bates, and others 1999; Petrides and Pandya 2002), which indicates that area 47/12 is strongly connected to brain regions in much the same way as observed in our data for the deep frontal operculum. The finding in subject III seems to support this, too: Here, the posterior portion of the orbital part of IFG is separated from the triangular part along the anterior branch of the lateral fissure and appears to have a very similar tractographic signature as the deep frontal operculum.
The parcellation of Broca's region exhibits a complex geometry and a large variability of the sulcal pattern (Ono and others 1990; Tomaiuolo and others 1999). The cytoarchitectonic studies of Amunts and others (2004) show that the borders of BA44 and BA45 principally correspond to this sulcal pattern; however, they also found frequent small and occasional larger deviations. The above described differences between connectivity-based area boundaries and the sulcal pattern are in line with this observation.
A way to correlate the regions identified on the basis of connectivity with cytoarchitectonic data is to compare them by means of probabilistic maps. We transformed the parcellation data from the individual subjects into the montreal neurological institute standard reference space (Collins and others 1994) and plotted the number of overlapping areas for each voxel of the standard brain. The resulting maps were then compared with the cytoarchitectonic probabilistic maps provided by the Institute of Medicine at the Research Centre in Juelich, Germany (http://www.fz-juelich.de/ime/ime_brain_mapping_eng) (see Fig. 7). Table 1 lists the mean centers of gravity of BA44 and BA45 in the cytoarchitectonic and connectivity-based probability maps. Generally, there is a principal agreement between the maps based on connectivity and cytoarchitecture: The areas based on connectivity are smaller and completely included within the respective cytoarchitectonic probability maps. Another noticeable feature of the connectivity-based maps is the smaller inferior–superior extent (especially BA45) and the more inferior location (BA44) of the area of maximal overlap (red in Fig. 7), respectively. Here, there might be a clue for some influence of macroanatomical differences between the 2 selections of brains. In the brains used in our study, the junction point between the precentral and inferior frontal sulci lies between z = 22 and z = 28 (95% confidence interval), which is considerably more inferior to the average values of z = 29 (Derrfuss and others 2004) and z = 30 (Germann and others 2005) found in literature. This might explain a reduced z-extent of the Broca area in our data. Generally, the comparison between the 2 probabilistic maps is hampered by the small sample sizes (10 subjects for the cytoarchitecture and 6 subjects for the tractographic parcellation), which render the probability maps less than representative.
|x||−42 (2)||−44 (2)||−42 (2)||−43 (2)|
|y||11 (7)||12 (1)||24 (7)||25 (1)|
|z||20 (4)||13 (5)||17 (5)||13 (2)|
|x||−42 (2)||−44 (2)||−42 (2)||−43 (2)|
|y||11 (7)||12 (1)||24 (7)||25 (1)|
|z||20 (4)||13 (5)||17 (5)||13 (2)|
The coordinates are given in standard stereotactic space (montreal neurological institute) with standard deviations (in parenthesis; n = 10 for cytoarchitectonic map; n = 6 for this study).
The interpretability of tractograms in terms of anatomical connectivity is vulnerable to artifacts and therefore to be taken with caution. Before any such interpretation can be attempted, a thorough methodological evaluation of the relationship between tractography and the actual anatomical connectivity should be performed. Therefore, we basically report our findings (see results) and just strive to discuss those aspects that compare with previous evidence from other methods in both human and nonhuman primates. In general, the tractographic signatures that we report above are largely in line with findings from monkey tracer studies (Pandya and Yeterian 1996; Romanski, Tian, and others 1999; Petrides and Pandya 2002) as well as with magnetic resonance imaging–based white matter parcellation in humans (Makris and others 1999; Mori and others 2005) and with other DTI inferences (e.g., Catani and others 2005). In accordance with literature, the dorsal pathway is thought to connect BA44 and (in our case to a somewhat lesser extent) BA45 with parietal and temporal association cortices (Parker and others 2005; Pulvermuller 2005). In particular, for putative BA45, the appearance of a dorsal and a middle pathway would be in line with a further distinction between an anterior and a posterior portion (BA45A/B, Petrides and Pandya 1994, 2002). We speculate that the dorsal pathway is attributed to BA45B—and that the connectivity pattern of this region would be more BA44 like—whereas the middle pathway rather connects BA45A with the anterior temporal lobe. It is interesting that the deep frontal operculum seems to be connected to structures of the limbic system. Therefore, hypothetically its anatomical connectivity appears to be similar to that of BA47(/12), but rather different from the surface portion of Broca's area, BA44, and BA45. This is particularly interesting in the light of Friederici and others finding that the deep frontal operculum may support the processing of a more primitive type of grammar, whereas the surface portion of Broca's area subserves the processing of a more complex grammar (Friederici and others 2006). Furthermore, the processing of local structures (local transitional probability) in language is known to involve the deep frontal operculum and the anterior portion of the superior temporal gyrus (Friederici and others 2003), whereas the processing of hierarchical structures, in contrast, is supported by an activation network involving BA 44, premotor cortex, posterior superior temporal gyrus and sulcus, and the intraparietal sulcus (Bornkessel and others 2005).
We have presented a method for the segregation of cortical areas in the individual living subject. This method is based on probabilistic tractography and relies on DW-MRI data. The replication of the previous results by Johansen-Berg and others (2004) provided confidence in this kind of parcellation technique. Using this method, we have segregated 3 regions of Broca's area. This segregation was found to be plausible in the light of previous microanatomical and functional findings. Putatively, the found regions represent BA44 and BA45 as well as the deep frontal operculum. We conclude that the proposed method for cortex parcellation seems capable of revealing aspects of the functio-anatomical organization of the cortex that are not possible to obtain by other methods in the individual living subject.
The authors wish to thank S. Kotz and J. Derfuss for fruitful discussions on our findings, H. Schmidt and S. Seifert for help with the figures and segmentation of the ROIs, respectively, as well as M. Brass for his part in the initiation of this project. Conflict of Interest: None declared.