White Matter Segmentation Based on a Skeletonized Atlas: Effects on Diffusion Tensor Imaging Studies of Regions of Interest

Abstract

Purpose:

To compare the influence of conventional and skeletonized atlas-based white matter (WM) segmentation on diffusion tensor imaging (DTI) region-of-interest (ROI) investigations.

Materials and Methods:

A conventional WM atlas was skeletonized by thinning the corresponding fractional anisotropy (FA) map and labels. The conventional and skeletonized versions of the atlas were used for WM segmentation. The percentage of non-WM voxels assigned to WM labels, as well as statistical summaries of tensor-derived quantities, were compared between segmentation approaches. The ability to detect small differences in diffusion properties across groups of subjects was also compared between segmentation approaches.

Results:

Skeletonized segmentation resulted in significantly lower non-WM percentage (P < 0.05), higher mean FA and lower trace (P < 0.05) in most WM labels, and mainly lower standard deviation of FA and trace in labels neighboring the ventricles. In terms of maximizing the ability to detect intergroup DTI differences, skeletonized segmentation was superior in the corpus callosum, but the optimal approach varied for other WM labels.

Conclusion:

Conventional and skeletonized atlas-based segmentation probe different portions of brain tissue and lead to different statistical summaries of diffusion characteristics in WM labels. Careful selection of segmentation approach is required for DTI investigations of WM ROIs.

IN DIFFUSION TENSOR imaging (1,2) (DTI) studies of brain white matter regions-of-interest (ROI), manual outlining of ROIs can be particularly time-consuming and may introduce user bias. Automated selection of white matter ROIs can be accomplished with atlas-based segmentation (3–7). According to this approach an image volume corresponding to a white matter atlas is registered to an individual’s image volume, and the resulting spatial transformation is applied to the white matter labels of the atlas to transform them to the individual’s space. However, the transformation of an atlas to an individual’s space may suffer from substantial inaccuracies, due to low quality in the image volume corresponding to the individual brain and/or the atlas, limitations of the registration algorithm, and structural brain differences across subjects due to pathology or normal anatomical variation (8). Due to the high contrast in diffusion tensor properties across brain white matter, gray matter, and cerebrospinal fluid, spatial mismatch of the transformed anatomical labels with an individual’s actual anatomy may have a substantial impact on the results of DTI ROI investigations.

Conventional voxel-wise DTI investigations also suffer from misregistration across subjects (8). It has been demonstrated that the effects of misregistration can be reduced in voxel-wise studies, by projecting information from voxels with the highest fractional anisotropy (2) (FA) onto a reference white matter skeleton (8) or tract surface (9), and conducting statistical analyses using only the projected information (8,9). This process is relatively immune to reasonable levels of misregistration across subjects. To benefit from this major advantage, several investigations on brain white matter ROIs have recently combined the concept of projection to a skeleton and that of atlas-based segmentation (10–17). More specifically, ROI characteristics were extracted through projection of information from individual subjects onto a single reference white matter skeleton with predefined anatomical labels (10–17), i.e., a skeletonized white matter atlas. However, the use of skeletonized atlas-based segmentation in studies of white matter ROIs has not yet been evaluated.

The purpose of this work was to compare the effects of conventional and skeletonized atlas-based segmentation on DTI investigations of white matter ROIs. Skeletonized atlas-based segmentation was shown to reduce partial volume effects caused by misregistration of individual datasets with the atlas. It was also shown that conventional and skeletonized atlas-based segmentation lead to different statistical summaries of the diffusion characteristics in the same white matter labels. Finally, it was demonstrated that, the ability to detect intergroup DTI differences in the corpus callosum is maximized with skeletonized segmentation, but for other white matter labels the preferred segmentation approach varies.

MATERIALS AND METHODS

Conventional and Skeletonized Atlas

The Eve white matter atlas was used in this work, because (a) it contains diffusion tensor information of the whole brain, thereby allowing tensor-based registration, (b) its labels provide substantial anatomical detail, (c) it is in digital format, and (d) it is freely available (3). The FA map corresponding to the conventional Eve atlas was thinned using the Tract-Based Spatial Statistics (8) (TBSS, Oxford, UK) software. This process generated the white matter skeleton corresponding to the conventional Eve atlas. The skeleton was then thresholded with FA > 0.25 to exclude voxels containing primarily gray matter or cerebrospinal fluid. Finally, the skeletonized version of the Eve atlas was generated by applying the white matter skeleton mask to the labels of the conventional Eve atlas (Fig. 1).

Figure 1.

Color maps of the white matter labels of the conventional (a) and skeletonized (b) Eve atlas, overlaid on the corresponding FA template (grayscale). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

MRI Data

Three human brain MRI datasets were used in this work. Dataset 1 was used to compare the percentage of non–white matter voxels assigned to each white matter label between segmentation approaches. Dataset 2 was used to compare statistical summaries of tensor-derived quantities between segmentation approaches. Dataset 3 was used to assess the ability of each of the two approaches to detect small intergroup differences in the diffusion properties of white matter. Written informed consent was provided by all participants according to procedures approved by the local institutional committees for the protection of human subjects.

Dataset 1 consisted of T1-weighted anatomical and DTI data from 20 healthy human subjects (6 male, 32.7 ± 5.1 years of age, 27–39 years of age, 14 female, 26.9 ± 5.2 years of age, 22–39 years of age), collected on a 3 Tesla (T) General Electric MRI scanner (GE, Waukesha, WI). T1-weighted data were acquired using a three-dimensional (3D) magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence with the following imaging parameters: echo time (TE) = 3.2 ms, repetition time (TR) = 3.1 s, preparation time = 725 ms, flip angle 6°, field of view 24 cm × 24 cm, 124 sagittal slices, 1.5 mm slice thickness, 192 × 256 k-space matrix reconstructed to 256 × 256. DTI data were acquired using a Turboprop-DTI (18) sequence with the following parameters: TR = 5800 ms, TE = 94 ms, 8 spinechoes per TR, 5 k-space lines per spin-echo, 128 samples per line, 16 blades per image, field-of-view 24 cm × 24 cm, 3 mm slice thickness, 36 axial slices, 256 × 256 image matrix, b = 900 s/mm² for 12 diffusion gradient directions uniformly distributed in 3D space (minimum energy scheme) (19), and two b = 0 s/mm² image volumes.

Dataset 2 consisted of DTI data from 20 healthy human subjects (11 male, 31.3 ± 4.1 years of age, 26–35 years of age, 9 female, 30.3 ± 3.6 years of age, 26–36 years of age), obtained from the IXI brain database (20) (http://www.brain-development.org). DTI data were collected on a 3T Philips MRI scanner (Best, Netherlands), using a single-shot echo-planar (EPI) DTI sequence with the following imaging parameters: TR = 12,000 ms, TE = 51 ms, field-of-view 224 mm × 224 mm, 2 mm slice thickness, 64 axial slices, 128 × 128 image matrix, b = 1000 s/mm² for 15 diffusion gradient directions, and parallel imaging with an acceleration factor of 2.

Dataset 3 consisted of DTI data from 22 healthy human subjects (12 male, 32.8 ± 4.2 years of age, 26–37 years of age, 10 female, 28 ± 3.6 years of age, 24–34 years of age), also obtained from the IXI brain database, and collected as described for Dataset 2.

Preprocessing

For all datasets, diffusion-weighted (DW) and b = 0 s/mm² image volumes were co-registered using TORTOISE (21) to correct bulk motion and eddy-current distortions. For Dataset 1, only bulk motion was corrected because Turboprop-DTI is relatively immune to eddy-currents and magnetic field nonuniformities (22). The b-matrix was reoriented and diffusion tensors were estimated using TORTOISE. Maps of FA and trace of the diffusion tensor were generated.

Conventional and Skeletonized Atlas-Based Segmentation

Conventional and skeletonized atlas-based segmentation was performed on all DTI data. For both segmentation approaches, the DTI data of each subject were registered to the diffusion tensor template of the Eve atlas using DTITK (23) for uniformity. For conventional atlas-based segmentation, the inverse spatial transformations were applied to the conventional Eve white matter labels to transform them to each subject’s raw DTI space. For skeletonized atlas-based segmentation, the DTITK-transformed FA information of each subject was projected to the Eve skeleton using TBSS (no TBSS registration was conducted on the DTITK-transformed FA image volumes). The skeletonized Eve white matter labels were then transformed to the subject’s raw DTI space, by first back-projecting the skeletonized labels to Eve space using TBSS, and applying the inverse DTITK transformations.

Non–White Matter DTI Signals Assigned to White Matter Labels

For each subject in Dataset 1, DTI signals not corresponding to white matter tissue were identified as follows. First, the T1-weighted MPRAGE data were segmented into white matter, gray matter, and cerebrospinal fluid using FreeSurfer (24). Voxels containing gray matter and cerebrospinal fluid were combined to generate a binary mask of non-white matter. The T1-weighted data were then registered to the corresponding b = 0 s/mm² image volume using rigid body registration and mutual information through DTITK (23). Rigid body registration is sufficient to spatially match data collected with MPRAGE and Turboprop-DTI (25). The resulting spatial transformation was then applied to the non–white matter mask to transform it to the space of raw DTI data. The same process was repeated for all subjects in Dataset 1.

For both segmentation approaches, the non–white matter mask was used to count the number of voxels that did not contain white matter tissue but were assigned to white matter labels. The percentage of non–white matter voxels assigned to each white matter label was calculated for each subject in Dataset 1 and each segmentation approach. A Monte Carlo permutation test was used in each label to test the hypothesis that, the non–white matter percentage was significantly lower for skeletonized compared with conventional atlas-based segmentation (10,000 permutations; differences were considered significant at P < 0.05). Maps of the difference in mean (over all subjects in Dataset 1) non–white matter percentage across segmentation approaches were generated.

White Matter Diffusion Characteristics

The diffusion characteristics extracted for each white matter label were compared across segmentation approaches. For each subject in Dataset 2, the mean FA and trace values were calculated in each white matter label, using both segmentation approaches. To assess if simple parametric regression and inference is valid for FA and trace values extracted with the two segmentation approaches, or if a permutation-based method is more appropriate, the Lilliefors (26) modification of the Kolmogorov–Smirnov test was used to identify labels where the distribution of FA or trace values across subjects was significantly non-Gaussian. Non-Gaussianity was assumed if more than 5% of the labels failed the test. Next, a Monte Carlo permutation test was used in each label to test for significant differences in FA and trace values across segmentation approaches (10,000 permutations; differences were considered significant at P < 0.05). Maps of the difference in mean and standard deviation (over all subjects in Dataset 2) of FA and trace across segmentation approaches were generated.

Ability to Detect Small Intergroup Differences in White Matter Diffusion Characteristics

The ability to detect small differences in white matter diffusion characteristics across groups of human subjects was investigated for both segmentation approaches. This was accomplished with a simulation, as well as with power analysis.

First, for each subject in Dataset 3, diffusion abnormalities were simulated throughout white matter (in voxels with FA > 0.25) by increasing the secondary and tertiary eigenvalues of the diffusion tensors by 3%, while maintaining all original eigenvectors and primary eigenvalues. The resulting data formed Dataset 3′. The spatial transformations and projections generated for Dataset 3 for the purposes of conventional and skeletonized atlas-based segmentation (described above) were also applied to Dataset 3′. Two groups of ten subjects each were randomly selected from Datasets 3 and 3′, without replacement (for each group). The mean FA and trace values were calculated in each white matter label, for each subject, using both segmentation approaches. A Monte Carlo permutation test was used in each label to test for significantly lower FA and higher trace values in the group drawn from Dataset 3′ compared with that drawn from Dataset 3 (10,000 permutations; differences were considered significant at P < 0.05), for each segmentation approach separately. The process of selecting two groups of subjects and comparing FA and trace values across groups for each segmentation approach was repeated 1000 times in a bootstrap manner. Maps of the number of times a significant difference was detected across groups were compared between segmentation methods.

The ability to detect small intergroup differences in diffusion characteristics was also assessed using power analysis. The mean and standard deviation of FA and trace values for each label were calculated across subjects in Dataset 3, for each segmentation approach separately. The number of subjects necessary to detect a 10% decrease in FA, or 10% increase in trace, was obtained using the G*Power software (27) assuming two groups with the same number of subjects, significance at P < 0.05, and power > 0.95. The sample size extracted from the power analysis was compared in each label across segmentation approaches.

RESULTS

Non–White Matter DTI Signals Assigned to White Matter Labels

The percentage of non–white matter voxels assigned to 88 of the 104 white matter labels (85%) was significantly lower when using skeletonized compared with conventional atlas-based segmentation (P < 0.05) (Table 1, Fig. 2). Significantly higher non–white matter percentage for skeletonized compared with conventional segmentation was detected in only one label (left column and body of the fornix, Fig. 2).

Table 1.

Non–White Matter Percentage^*

White matter label	Skeletonized segmentation (%)	Conventional segmentation (%)	P value
Right uncinate fasciculus	5.7 ± 7.0	55.4 ± 11.9	<10−4
Left fusiform white matter	1.7 ± 7.5	29.9 ± 15.0	<10−4
Left uncinate fasciculus	24.1 ± 10.4	51.2 ± 10.1	<10−4
Right lateral fronto-orbital white matter	0.8 ± 1.2	23.4 ± 6.3	<10−4
Right middle fronto-orbital white matter	5.5 ± 7.5	27.9 ± 10.8	<10−4
Right fornix (column and body)	26.2 ± 21.0	45.1 ± 14.1	<10−4
Left cingulum (hippocampus)	4.8 ± 19.3	22.2 ± 16.3	<10−4
Left middle fronto-orbital white matter	5.6 ± 8.3	20.6 ± 9.5	<10−4
Left lateral fronto-orbital white matter	3.9 ± 3.7	18.9 ± 10.1	<10−4
Right inferior fronto-occipital fasciculus	5.7 ± 3.2	19.6 ± 5.8	<10−4

^*The non-white matter percentage for 88 of the 104 white matter labels was significantly lower when using skeletonized compared to conventional atlas-based segmentation. This table presents the mean and standard deviation of the non–white matter percentage for 10 of these labels that have the highest difference across segmentation approaches.

Figure 2.

Color maps of the difference in mean non–white matter percentage in Dataset 1 across segmentation approaches (skeletonized - conventional). Cool colors indicate a higher non–white matter percentage for conventional segmentation, and warm colors indicate a higher non–white matter percentage for skeletonized segmentation. Only labels with significant differences are included in these maps. T2-weighted images of the atlas are shown in the background in grayscale. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

White Matter Diffusion Characteristics

When using skeletonized atlas-based segmentation, the percentages of labels for which FA or trace values did not follow a Gaussian distribution were 2.9% and 8.7%, respectively. When using conventional atlas-based segmentation, the percentages of labels for which FA or trace values did not follow a Gaussian distribution were 7.7% and 17.3%, respectively.

FA values in 100 of the 104 white matter labels (96%) were significantly higher when using skeletonized compared with conventional atlas-based segmentation (P < 0.05) (Table 2) (Fig. 3a). Significantly lower FA values when using skeletonized compared with conventional atlas-based segmentation were not detected in any white matter labels. Trace values in 86 of the 104 white matter labels (83%) were significantly lower when using skeletonized compared with conventional atlas-based segmentation (P < 0.05) (Table 3, Fig. 4a). Significantly higher trace values when using skeletonized compared with conventional atlas-based segmentation (P < 0.05) were detected in only 3 of the 104 white matter labels (2%) (left column and body of the fornix, retrolenticular part of the left internal capsule, right cingulum [cingulate gyrus]) (Fig. 4a). The standard deviation of FA and trace across subjects was mainly lower in labels neighboring the ventricles when using skeletonized compared with conventional segmentation (Figs. 3b, 4b). In the rest of the brain, the relation between the standard deviation of FA and trace obtained with the two segmentation approaches varied (Figs. 3b, 4b). Figures 3 and 4 show differences in mean and standard deviation of FA and trace in Dataset 2 across segmentation approaches, for labels with significant differences in mean values.

Table 2.

FA Values^*

White matter label	Skeletonized segmentation	Conventional segmentation	P value
Left fusiform white matter	0.60 ± 0.09	0.32 ± 0.08	9.6×10−3
Left pons	0.68 ± 0.06	0.46 ± 0.06	<10−4
Left cingulum (hippocampus)	0.67 ± 0.04	0.48 ± 0.03	<10−4
Right fornix (column and body)	0.72 ± 0.05	0.53 ± 0.03	<10−4
Right body of corpus callosum	0.72 ± 0.02	0.55 ± 0.04	<10−4
Right uncinate fasciculus	0.54 ± 0.07	0.37 ± 0.04	<10−4
Right cingulum (hippocampus)	0.62 ± 0.04	0.46 ± 0.03	<10−4
Left tapetum	0.62 ± 0.05	0.46 ± 0.06	<10−4
Right inferior cerebellar peduncle	0.68 ± 0.03	0.52 ± 0.02	<10−4
Left cingulum (cingulate gyrus)	0.62 ± 0.03	0.47 ± 0.03	<10−4

^*FA values in 100 of the 104 white matter labels significantly higher when using skeletonized compared to conventional atlas-based segmentation. This table presents the mean and standard deviation of FA for 10 of these labels that have the highest difference across segmentation approaches.

Figure 3.

Color maps of the difference in mean (a) and standard deviation (b) of FA in Dataset 2 across segmentation approaches (skeletonized - conventional). Cool colors indicate a higher mean or standard deviation of FA for conventional segmentation, and warm colors indicate a higher mean or standard deviation of FA for skeletonized segmentation. Only labels with significant FA differences are included in these maps. T2-weighted images of the atlas are shown in the background in grayscale. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Table 3.

Trace Values^*

White matter label	Skeletonized segmentation (10−3 mm2/s)	Conventional segmentation (10−3 mm2/s)	P value
Left tapetum	2.36 ± 0.13	3.65 ± 0.67	<10−4
Right fornix (column and body)	2.48 ± 0.31	3.60 ± 0.23	<10−4
Right body of corpus callosum	2.35 ± 0.10	3.27 ± 0.36	<10−4
Right tapetum	2.40 ± 0.13	3.07 ± 0.64	<10−4
Right cuneus white matter	2.01 ± 0.07	2.63 ± 0.67	<10−4
Left body of corpus callosum	2.33 ± 0.10	2.91 ± 0.25	<10−4
Right splenium of corpus callosum	2.14 ± 0.07	2.69 ± 0.22	<10−4
Right cerebral peduncle	2.16 ± 0.07	2.63 ± 0.14	<10−4
Left medulla	2.22 ± 0.47	2.67 ± 0.24	<10−4
Left pons	1.91 ± 0.17	2.33 ± 0.15	<10−4

^*Trace values in 86 of the 104 white matter labels significantly lower when using skeletonized compared to conventional atlas-based segmentation. This table presents the mean and standard deviation of trace values for 10 of these labels that have the highest difference across segmentation approaches.

Figure 4.

Color maps of the difference in mean (a) and standard deviation of trace in (b) Dataset 2 across segmentation approaches (skeletonized - conventional). Cool colors indicate a higher mean or standard deviation of trace for conventional segmentation, and warm colors indicate a higher mean or standard deviation of trace for skeletonized segmentation. Only labels with significant trace differences are included in these maps. T2-weighted images of the atlas are shown in the background in grayscale. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Ability to Detect Small Intergroup Differences in White Matter Diffusion Characteristics

In the simulation, for 40 white matter labels (38%), skeletonized segmentation led to detection of a significant FA difference across groups in a higher number of bootstraps than conventional segmentation (Fig. 5a). For 61 labels (59%), conventional segmentation led to detection of significant FA differences across groups in more bootstraps than skeletonized segmentation (Fig. 5b). Also in the simulation, for 55 white matter labels (53%), skeletonized segmentation led to detection of a significant trace difference across groups in a higher number of bootstraps than conventional segmentation (Fig. 5c). For 48 white matter labels (46%), conventional segmentation led to detection of significant trace differences across groups in more bootstraps than skeletonized segmentation (Fig. 5d).

Figure 5.

a,c: Color maps of the labels in which skeletonized segmentation allowed detection of significant FA (a) or trace differences (c) in a higher number of bootstraps than conventional segmentation. Different colors correspond to different values of the ratio of the number of bootstraps with significant FA (a) or trace differences (c) for skeletonized over conventional segmentation. b,d: Color maps of the labels in which conventional segmentation allowed detection of significant FA (b) or trace differences (d) in a higher number of bootstraps than skeletonized segmentation. Different colors correspond to different values of the ratio of the number of bootstraps with significant FA (b) or trace differences (d) for conventional over skeletonized segmentation. Only labels with ratios higher than one are displayed. T2-weighted images of the atlas are shown in the background in grayscale.

Power analysis showed that, for 24 white matter labels (23%), skeletonized segmentation required a higher sample size than conventional segmentation to detect a 10% decrease in FA (Fig. 6a). For 60 labels (58%), conventional segmentation required a higher sample size than skeletonized segmentation to detect a 10% decrease in FA (Fig. 6b). Additionally, power analysis demonstrated that, for 30 white matter labels (29%), skeletonized segmentation required a higher sample size than conventional segmentation to detect a 10% increase in trace (Fig. 6c). For 34 labels (33%), conventional segmentation required a higher sample size than skeletonized segmentation to detect a 10% increase in trace (Fig. 6d).

Figure 6.

a,c: Color maps of the labels in which skeletonized segmentation required a higher sample size than conventional segmentation to detect a 10% decrease in FA (a), or 10% increase in trace (c). Different colors correspond to different values of the ratio of sample sizes for skeletonized over conventional segmentation. b,d: Color maps of the labels in which conventional segmentation required a higher sample size than skeletonized segmentation to detect a 10% decrease in FA (b), or (d) 10% increase in trace. Different colors correspond to different values of the ratio of sample sizes for conventional over skeletonized segmentation. Only labels with ratios higher than one are displayed. T2-weighted images of the atlas are shown in the background in grayscale.

DISCUSSION

Several DTI investigations have recently used a skeletonized atlas for white matter ROI selection (10–17) to reduce misregistration effects of conventional atlas-based segmentation on the final results. The purpose of this work was to compare the influence of conventional and skeletonized atlas-based segmentation on DTI investigations of white matter ROIs. As anticipated, skeletonized segmentation was shown to reduce partial volume effects caused by misregistration of individual datasets with the atlas. Furthermore, it was demonstrated that the two segmentation approaches led to different statistical summaries of the diffusion characteristics of white matter labels. Finally, although skeletonization reduces partial volume effects, the segmentation approach that maximizes the ability to detect small intergroup DTI differences varied throughout white matter. The results of the present study are discussed in detail below.

The high number of white matter labels with a significantly lower percentage of non–white matter voxels when using skeletonized compared with conventional atlas-based segmentation was anticipated and was due to the fact that, in conventional segmentation, misregistration between individual datasets and the atlas leads to partial volume effects with gray matter and/or cerebrospinal fluid. In contrast, projection of information onto a white matter skeleton makes skeletonized segmentation relatively immune to reasonable amounts of misregistration (on the order of a few mm, also shown in Smith et al, (8)). The quality of the individual DTI data influences registration accuracy to the atlas and thereby the exact percentages of non–white matter voxels. However, the main conclusion of this analysis is expected to remain true for any DTI acquisition technique and typical data quality.

For some white matter labels, the percentage of non–white matter voxels was substantially higher than zero for skeletonized segmentation. Furthermore, in one label the non–white matter percentage was significantly higher in skeletonized compared with conventional segmentation. These latter findings were due to imperfections in the non–white matter tissue masks, FA noise, and the combination of these factors with the small size of the regions in discussion.

For both segmentation approaches, registration of individual DTI data to the atlas was accomplished using DTITK (23), which is shown to perform at a high level compared with other published DTI registration algorithms (28). Also, an atlas from the same age group as the individual subjects was used throughout this study. When less effective registration techniques and/or less representative atlases are used for conventional atlas-based segmentation, the spatial matching deteriorates (29) and the percentage of non–white matter voxels assigned to white matter labels is expected to increase. For skeletonized segmentation, the percentage of non–white matter voxels is expected to remain relatively immune to increasing levels of misregistration, because the information projected onto the skeleton originates from voxels with high FA. Instead, skeletonized segmentation under severe misregistration will assign white matter DTI information to the wrong labels. For skeletonized segmentation, a pronounced increase in the percentage of non–white matter voxels is expected to occur for data with high FA values in voxels not containing white matter, as is the case for data with low signal to noise ratio (30).

FA values extracted with skeletonized segmentation were shown to be non-Gaussian distributed in 2.9% of the labels, and with conventional segmentation in 7.7% of the labels. Because non-Gaussianity was expected to occur by chance in 5% of the labels, simple parametric regression and inference may be valid only for FA values extracted with skeletonized segmentation (note that the FA itself is inherently non-Gaussian due to its derivation). A permutation-based method is more appropriate for statistical analysis of FA values generated from conventional segmentation, as well as for trace values generated from both approaches (non-Gaussianity of trace values was demonstrated in 8.7% of the labels in skeletonized and 17.3% of the labels in conventional segmentation).

The fact that, in the overwhelming majority of white matter labels (96%), FA values extracted with skeletonized segmentation were significantly higher than those extracted with conventional segmentation was anticipated and was due to two reasons. First, inherent to the process of skeletonization is the projection onto the skeleton of DTI information originating from high FA voxels located centrally in white matter tracts. In contrast, conventional segmentation aims to summarize information from the whole width of white matter structures, including not only the high FA voxels selected during skeletonization, but also voxels near the edges of white matter structures that are typically characterized by lower FA values. Second, FA values in non–white matter tissue are generally lower than those in white matter, and the percentage of non–white matter voxels was shown to be higher for conventional compared with skeletonized segmentation. For these reasons, FA values in the overwhelming majority of white matter labels were significantly higher when using skeletonized compared with conventional segmentation.

Trace values extracted with skeletonized segmentation were significantly lower than those extracted with conventional segmentation in the majority of white matter labels (83%), due to the higher trace values in gray matter and cerebrospinal fluid compared with white matter, combined with the higher percentage of non–white matter voxels for conventional compared with skeletonized segmentation. Trace values were particularly higher for conventional compared with skeletonized segmentation for white matter tissue near large bodies of cerebrospinal fluid. Finally, the unexpected finding of trace values that were statistically higher for skeletonized segmentation in 3 labels, was probably due to the small size of two of the regions and FA noise (per voxel) leading to projection errors.

The significant differences in mean FA and trace values for white matter labels segmented with the two approaches, are a manifestation of the fact that the two types of segmentation extract information from different segments of brain tissue. Skeletonized segmentation measures almost exclusively the characteristics of the central skeleton of white matter, while conventional segmentation probes more of the white matter, and may be contaminated from gray matter and cerebrospinal fluid. Furthermore, factors such as brain atrophy or lesions may affect the measurements in conventional but not necessarily skeletonized segmentation. Consequently, the relation of tensor-derived metrics with demographic or clinical variables may not necessarily be the same for the two types of segmentation. Therefore, attention is required when evaluating and comparing the results of studies using different segmentation approaches.

Although skeletonized segmentation reduced partial volume effects in the vast majority of white matter labels, there was a substantial number of labels for which conventional segmentation provided more sensitive detection of small diffusion changes. This was probably due to the effects of noise in combination with the drastically lower number of voxels assigned to labels in skeletonized compared with conventional segmentation. Therefore, care should be taken when selecting segmentation method for detecting intergroup DTI differences (10–14). The corpus callosum and white matter labels directly neighboring the ventricles appeared to substantially benefit from the use of skeletonized segmentation. In the rest of the brain, skeletonized and conventional segmentation both improved detection of DTI group differences in different regions. For different white matter labels than the ones used here (different atlas), the segmentation approach that enhances the ability to detect intergroup DTI differences must be reevaluated.

The present study has certain limitations. First, in the simulation of DTI data with white matter abnormalities, the diffusion characteristics were altered in all voxels with an original FA > 0.25. Therefore, the simulation covered only situations where the white matter changes are rather diffuse. The two segmentation approaches were not compared for the case of focal abnormalities. Depending on the size, location, and properties of such foci, skeletonized segmentation may not even sample information from the affected voxels. Therefore, skeletonized segmentation may not be appropriate for studies of conditions involving focal abnormalities. In general, the purpose of the simulation was to provide an example of the differences between the two atlas-based segmentation approaches, and not to demonstrate their performance for all possible types and spatial patterns of white matter abnormalities. Second, automated selection of white matter ROIs can also be accomplished with fiber tractography based on a set of carefully defined seeds (31). This approach was not compared with atlas-based segmentation in the present work. However, tractography results depend heavily on the image quality of the individual diffusion datasets and disorders that may be present. Additionally, user-dependent manual editing of the results is often necessary to reject erroneous tracts. Furthermore, selected ROIs include whole fiber bundles often consisting of voxels with very different diffusion properties, instead of smaller more uniform white matter segments. Also, it has been demonstrated that the tensor model is not ideal for tractography purposes (32). Third, the present work focused on quantities derived from the diffusion tensor model, and caution should be exercised when interpreting the results for studies based on alternative diffusion models.

In conclusion, this work compared the effects of conventional and skeletonized atlas-based segmentation on DTI investigations of white matter ROIs. Skeletonized segmentation was shown to reduce partial volume effects. Additionally, it was demonstrated that the two segmentation approaches generate different statistical summaries of the diffusion characteristics of white matter labels, because they probe different segments of brain tissue. Finally, it was shown that the segmentation approach that maximizes the ability to detect small intergroup DTI differences varies throughout white matter. Therefore, caution should be exercised when selecting segmentation approach for DTI investigations of white matter ROIs.