In the present study, we compared an atlas-based approach, a streamline tractography analysis approach, and a combined approach applied to the same DTI data to examine the relative accuracy of each method for predicting outcome in children with ABI, and to evaluate the advantages and disadvantages of each method (and their pre-processing approaches) for possible clinical use. We investigated the agreement between each method and the accuracy of each method for predicting clinical outcome using FA data from the CST.
For the primary aim of this study, we found differences between the FA values derived using the different approaches. For the second aim of this study, we found that the FA values from the ipsilesional CST calculated using both approaches (and the NT from streamline tractography) are significantly associated with the motor outcome after rehabilitation, but the atlas-based approach and combined approach incorporating an extra motion correction step had higher predictive accuracy for motor outcome, as assessed by ROC analysis.
One advantage of both streamline tractography and probabilistic tractography methods over atlas-based approaches is that they can be applied in native space, without registration to a template. While image registration and normalisation methods have demonstrated high reliability and accuracy in healthy brains [16], registration methods can provide inappropriate solutions when applied to images from patients with stroke or other brain injuries, where lesioned brains are matched to a template with no brain lesions [17, 18]. However, while atlas-based analyses can be automated to a large extent, streamline tractography results are dependent on manual delineations of the seed and target regions, so the results are arguably more operator-dependent than those from atlas-based analyses. In addition, streamline tractography methods are less accurate in cases where the raw DTI images were collected with a non-isotropic voxel resolution. Clinical DTI protocols like those utilised in the present study with a slice thickness (3 mm) considerably greater than the reconstructed in-plane resolution (0.94 × 0.94 mm) may, therefore, be less suited to streamline tractography analyses than protocols with an isotropic voxel resolution [19].
Operator dependence is also an important consideration for (non-tractographic) ROI analyses of DTI data. For example, Lilja et al. [20] showed that in quantitative DTI analyses of the optical tracts, results differ according to which ROI method (manual or semi-automatic) is applied. Similarly, Foeling et al. [21] reported high operator dependence associated with ROI analyses, after considering the importance of various factors like the ROI definition, atlas-based analyses, effects of motion, registration, and spatial normalisation. In a comparison of voxel-based and manual ROI-based analyses of DTI data in children and young adults, Snook et al. [22] observed good correlation between the FA values derived with automated and manual methods. However, they noted differences between the two methods in sensitivity to age effects in certain brain regions, thought to be due to the effects of spatial normalisation and smoothing in the voxel-based analyses. Based on the apparent differences in results between the methods they concluded that both manual (ROI-based) and voxel-based analyses offer complementary insight into neurodevelopment.
In the present study, the FA values derived with the streamline tractography analysis were significantly higher than those from the atlas-based analysis, despite moderate correlation between these two measures. The lower FA values measured with the atlas-based approach may be due to the inclusion of parts of the CST where degeneration reduced the FA beneath the tracking threshold for streamline tractography. Alternatively, these lower FA values may be due to partial volume effects. In the case of ABI, it is difficult to separate between these two effects (of degeneration and partial volume), as degeneration of the tract would be expected to cause a loss of white matter volume and a reduction in the “number of tracts” in the region previously occupied by the tract prior to the injury. However, the higher predictive accuracy for outcome observed with the FA values from the atlas-based approach suggests that the atlas-based FA values are clinically relevant, even if the atlas region includes parts of the tract that have undergone degeneration, or where the FA falls below the tracking threshold.
Our data also showed a link between the NT from streamline tractography and the motor outcome after rehabilitation. However, since the NT is vulnerable to bias from both experimental and biological factors [23], the FA may be a more robust indicator of white matter integrity, particularly in the clinical setting where the signal-to-noise ratio may be suboptimal and experimental parameters (including the voxel resolution) may vary. The FA values from the atlas-based method also seem to capture both aspects of degeneration (loss of volume/tracts and a reduction of FA in the remaining tracts) in a single measure, while these aspects are quantified separately with the NT and FA from the streamline tractography approach. While some previous studies have reported dependence of FA on the TE of the DTI acquisition [24], in the present study the link between FA and outcome remained unchanged after controlling for TE, suggesting that TE variations are unlikely to bias the apparent link between FA and outcome.
The combined approach (which used identical pre-processing to the streamline tractography method, including a motion correction step and eddy current correction, tensor fitting, and calculation of the FA maps) improved the prediction and was comparably or slightly more accurate for outcome in comparison to the atlas-based approach. Motion correction could most likely bring additional improvement to the accuracy of DTI for outcome prediction in datasets demonstrating significant motion during the scan.
Although the FA in the ipsilesional CST seems to provide a robust predictor of motor outcome, by considering only a single tract it is not possible to confirm the specificity of these findings, or whether, for example, FA values in other tracts also might predict motor outcome. Therefore, in order to test the regional specificity of the findings, we also exported the mean FA for the contralesional CST using the combined approach, and repeated the nonparametric partial correlation testing of the contralesional FA versus outcome. This additional analysis was performed just in the subgroup of patients with stroke, since some of the patients with TBI demonstrated bilateral injuries, potentially affecting both the contralesional and ipsilesional motor tracts. We found significant correlation between the WeeFIM motor outcome and the ipsilesional FA in the stroke subgroup, which was not present in the contralesional CST, providing some support for the regional specificity of the ipsilesional CST for motor outcome.
Several limitations should be taken into account for this study. As discussed by Soares et al. [6], there are many different software packages and tools available to pre-process and analyse DTI data. In this study, we only compared two particular implementations (an atlas-based method using FSL and a streamline tractography method using ExploreDTI), which both included pre-processing, tensor estimation, and tract selection. However, other methods (e.g., manual ROI analysis, voxel-based analyses, and probabilistic tractography) and software packages (e.g., SPM, Freesurfer, BrainVoyager, DoDTI, DTIstudio, Camino, etc.) are available for DTI data analyses, which were not considered in the present study. In addition to the FA, there are also other DTI metrics like the mean diffusivity, axial diffusivity, and radial diffusivity [25], which were not considered. Further comparisons between analysis methods and software implementations and across DTI metrics would be needed to establish the relative advantages and disadvantages of each approach, and the relative sensitivity of each DTI metric to outcome.
Another limiting factor is the heterogeneity of the patient group and the relatively small sample size, which made it difficult to perform an analysis like the ROC analysis for each subgroup (stroke versus TBI) or each DTI protocol (21 versus 35 directions). For example, in the group of patients in whom the DTI was acquired with 35 directions, only three patients had a poor outcome, and within the TBI group only three patients had a poor outcome. The timing of the MRI measurement after the injury also varied across the patient group. Some patients were measured first in another hospital using another MRI scanner (1.5-T field strength, not used in this study), or with computed tomography. Nevertheless, all MRI data included in the present study were acquired before the rehabilitation therapy and the time period between injury and MRI was included as a covariate in the correlation analysis. This group variability could not be corrected due to the retrospective nature of this study, but future studies incorporating larger patient group sizes, or with a prospective design may be able to account for variability in outcome arising from differences in the type or timing of acquired brain injury, or in the scanning protocol.
In conclusion, for the primary aim of this study, we found differences between the FA values derived using the different approaches. For the second aim, in a clinical DTI sample of children with ABI, FA values from streamline tractography were higher than those from the atlas-based and the combined approach. FA values for the CST derived from an atlas-based approach and combined approach provide better predictive accuracy for clinical outcome than those derived from streamline tractography. Nevertheless, FA values from both methods provide significant predictors for clinical motor outcome. The combined approach utilising an additional motion correction step seems to improve the accuracy of DTI as a predictor of the rehabilitation outcome.