Effect of CSF Suppression for Diffusional Kurtosis Imaging

Abstract

Purpose

To evaluate the cerebrospinal fluid (CSF) partial volume effect on diffusional kurtosis imaging (DKI) metrics in white matter and cortical gray matter.

Materials and Methods

Four healthy volunteers participated in this study. Standard DKI and fluid-attenuated inversion recovery (FLAIR) DKI experiments were performed using a twice-refocused-spin-echo diffusion sequence. The conventional diffusional tensor imaging (DTI) metrics of fractional anisotropy (FA), mean, axial, and radial diffusivity (MD, D_∥, D_⊥) together with DKI metrics of mean, axial and radial kurtosis (MK, K_∥, K_⊥) were measured and compared. Single image slices located above the lateral ventricles, with similar anatomical features for each subject, were selected to minimize the effect of CSF from the ventricles.

Results

In white matter, differences of less than 10% were observed between diffusion metrics measured with standard DKI and FLAIR-DKI sequences, suggesting minimal CSF contamination. For gray matter, conventional DTI metrics differed by 19% to 52%, reflecting significant CSF partial volume effect. Kurtosis metrics, however, changed by 11% or less, indicating greater robustness with respect to CSF contamination.

Conclusion

Kurtosis metrics are less sensitive to CSF partial voluming in cortical gray matter than conventional diffusion metrics. The kurtosis metrics may then be more specific indicators of changes in tissue microstructure, provided the effect sizes for the changes are comparable.

INTRODUCTION

Diffusion tensor imaging (DTI) has been widely used to study diffusion properties of water in human brain (1–6). Using DTI, the fractional anisotropy (FA), as well as the mean, axial, and radial diffusivities (MD, D_∥, D_⊥) can be calculated (6,7), providing valuable information about the physical environment of water in brain tissue. One difficulty in the accurate quantification of these parameters, however, is the well-known cerebral spinal fluid (CSF) partial volume effect (8,9).

Previous studies have investigated this CSF partial volume effect by applying fluid-attenuated inversion recovery (FLAIR) prior to diffusion-weighted imaging to suppress the CSF contamination (10–17). These studies have found that CSF suppression with FLAIR can significantly reduce measured diffusivities and increase measured FA values, particularly in cortical gray matter and parenchymal tissue bordering the ventricles, indicating CSF contamination to be a potentially important confounding effect for DTI measurements in brain.

A recently introduced diffusion MRI technique called diffusional kurtosis imaging (DKI) extends DTI by simultaneously measuring the non-Gaussian behavior of water diffusion, in addition to the conventional DTI parameters (18–22). From a DKI dataset, one obtains metrics related to the kurtosis of the diffusion displacement probability distribution, including the mean, axial, and radial kurtoses (MK, K_∥, K_⊥). The kurtosis is a general dimensionless statistic that quantifies the non-Gaussianity of arbitrary distribution functions (23). Non-Gaussian diffusion is believed to arise from diffusion barriers, such as cell membranes and organelles, and water compartments (e.g., extracellular and intracellular). Therefore, the DKI metrics can be considered as natural indicators of tissue microstructural complexity in both gray and white matter structures.

Preliminary work suggests that the additional information provided by DKI could be useful in the study of several types of brain pathologies, including stroke (24–26), cancer (27,28), trauma (29–31), Huntington's disease (32), and attention-deficit hyperactivity disorder (33). In addition, the application of DKI in place of DTI may improve the accuracy of estimates for the conventional diffusion tensor parameters (34). Consequently, it is of interest to examine the degree to which CSF partial voluming influences diffusion metrics measured by DKI and to assess the extent CSF contamination can be a confounding effect for their interpretation.

We note that the effect of CSF contamination on the measured kurtosis of a brain tissue voxel is not a priori obvious. Pure CSF has an intrinsically low kurtosis (slightly elevated from zero due to flow effects), which might suggest that a CSF admixture would reduce the kurtosis. On the other hand, diffusional heterogeneity tends to increase the kurtosis (20), and thus a CSF component could plausibly raise the measured value.

In analogy with prior DTI investigations (11–15,17), the goal of the present study is to quantify CSF partial volume effects for DKI by comparing, in healthy human volunteers, the values for diffusion metrics obtained using standard DKI and FLAIR-DKI. The metrics considered here are the mean, axial, and radial diffusivities, the fractional anisotropy, and the mean, axial, and radial kurtoses. The relative changes in these parameters due to CSF suppression from the use of FLAIR provide indicators for the degree of CSF contamination in the standard DKI measurements. While a brief preliminary analysis of our data has been previously published as a conference abstract (35), that work did not give results for the axial and radial diffusion metrics and used a less highly optimized post-processing algorithm.

MATERIALS AND METHODS

Subjects

Four healthy volunteers ranging from ages 22 to 51 years were examined. Subjects were recruited from the local community of our institution. The protocol was approved by the NYU School of Medicine Institutional Review Board, and all subjects gave informed written consent before participating in the study.

MR imaging

MR experiments were conducted on a 3 Tesla (T) TIM Trio MR system (Siemens Medical Solutions, Erlangen, Germany) using a body coil for RF transmission and an 8-element phased array coil for signal reception. The DKI pulse sequence and processing algorithm have been described previously (19–21,36). Briefly, this technique uses a diffusion-sensitizing pulse sequence and acquires three or more b-values (in contrast to two b-values for conventional DTI) to evaluate the non-linearity of the log-signal decay, ln(S), as a function of diffusion weighting, b. (Note that, in DTI theory, ln(S) is assumed to be a linear function of b.) The experimental data is fitted to a quadratic function, in which the coefficients associated with the linear term give the apparent diffusion coefficient (ADC) and the coefficients associated with the quadratic term give the apparent diffusional kurtosis (ADK). In the special case of free diffusion (such as in a water bottle), the ADK will be zero and the model reduces to the DTI equation. Multiple gradient encoding directions can be used to obtain direction-dependent ADC and ADK values, from which one can calculate the diffusion tensor (DT) and the diffusional kurtosis tensor (DKT) (19–21,36). From the DT, one may derive several conventional diffusion metrics including the MD, D_∥, D_⊥ and the FA. From the DKT, one may obtain additional diffusion metrics, including the MK, K_∥, and K_⊥.

Standard DKI and FLAIR-DKI experiments were performed using a twice-refocused-spin-echo diffusion sequence (37). This sequence reduces the eddy-current-related distortions in the diffusion-weighted images. A total of thirty different diffusion encoding directions were used. The vectors for the directions were adapted from an optimized sampling strategy reported in the literature (38, 39). For each direction, six b-values (b = 0, 500, 1000, 1500, 2000, 2500 s/mm²) were employed. Other imaging parameters were: field of view (FOV) = 256×256 mm², acquisition matrix = 128×128, parallel imaging factor of 2 with 24 k-lines used as references, number of averages = 2, 15 anterior commissure/posterior commissure aligned slices to cover the frontal and temporal brain regions, slice thickness = 4 mm, interslice gap = 1 mm, voxel size = 2×2×4 mm³, repetition time (TR)/inversion time (TI)/echo time (TE) = 6800/2230/108 ms for FLAIR-DKI and TR/TE = 2300/108 ms for standard DKI. The scan durations were 35 min for FLAIR-DKI and 11 min and 57 sec for standard DKI.

Data processing

DKI parametric maps were obtained using software developed based on methods described elsewhere (36). Briefly, motion correction was performed on the diffusion-weighted images, followed by averaging and spatial smoothing of averaged images using a Gaussian kernel with a full width at half maximum of 3.375 mm. At each voxel, the diffusion tensor and the kurtosis tensor were simultaneously fitted to ln(S) using a constrained weighted linear least squares algorithm. The diffusion metrics were then estimated using tensor-based closed-formed expressions (20,36). The above steps were applied to both standard and FLAIR-DKI data to obtain seven individual parametric maps for each imaging method.

A slice just above the lateral ventricles with similar anatomical features was selected for each subject to minimize the effect of CSF from the ventricles. From prior studies (9–12,15), one may expect tissues bordering the ventricles to be strongly influenced by CSF partial voluming, but with a significant dependence on the precise choice of regions of interest (ROI). Here we decided to focus on the more subtle effects due to CSF in sulci, which are simpler to quantify in a meaningful way. The images obtained without diffusion weighting (i.e., b = 0) were skull-stripped automatically by using FMRIB Software Library (FSL) (University of Oxford, Oxford, UK). White matter and gray matter masks were then derived from the standard DKI maps and applied to both the standard and FLAIR-DKI parametric maps for all of the diffusion metrics. In order to help assess the robustness of our results, gray matter and white matter were segmented in two distinct ways. The first set of masks was based on an FA threshold of 0.15, and the second set of masks was based on an MK threshold of 1.0. For both types of mask, voxels with values (i.e., FA or MK) below the threshold were classified as gray matter, and voxels with values above the threshold were classified as white matter. The FA threshold corresponds to that utilized in the prior FLAIR-DTI study by Ma et al. (14) and has also been used by Stieltjes et al. (40) in the context of white matter fiber tracking. The threshold of MK = 1.0 was applied with the chosen value based on work of Falangola et al. (41). A potential advantage of using the MK for segmentation rather than the FA is that it is less affected by fiber crossings. For both cases, gray and white matter masks were generated using the standard DKI maps for each subject and then used to delineate ROI for all of the subject's parametric maps. Mean values of all the voxels for the ROI were obtained with ImageJ (National Institutes of Health, Bethesda, MD, USA). These values were averaged across all four subjects, and the differences between values obtained from standard DKI and FLAIR-DKI were compared.

Percent changes in the diffusion metrics were calculated according to

$$\text{Percent change}\equiv 100\times \frac{X_{\text{standard}}-X_{\text{FLAIR}}}{X_{\text{FLAIR}}},$$ [1]

where X_standard represents any of the diffusion metrics as measured with standard DKI and X_FLAIR represents the corresponding metric as measured with FLAIR-DKI. We chose the convention with X_FLAIR in the denominator, as we regard the values with CSF suppression to be better estimates for the “true” parenchymal diffusion metrics.

To investigate the effect of explicit CSF removal by means of segmentation, CSF masks were derived from the b = 0 images obtained without inversion recovery. The diffusivity of CSF is near that of pure water at 37°C, which is about 3.0 μm²/ms (42). Therefore voxels with MD values exceeding 2.5 μm²/ms were classified as being substantially CSF. The comparison of the standard and FLAIR-DKI diffusion metrics were then repeated as above, but with all the CSF dominated voxels eliminated from the analysis by means of the CSF masks. Both the FA-based and MK-based gray/white matter segmentations were considered.

In order to quantify the signal-to-noise ratio (SNR) for the b = 0 images, square ROI consisting of 49 voxels were defined in white matter and in air (upper right corner) for both the standard and FLAIR acquisitions. The SNR was then defined as the ratio of the mean signal in white matter to that of air. While there are other, arguably more sophisticated methods of computing SNR, this simple approach was deemed adequate for assessing the relative SNR of the standard and FLAIR images (43). White matter was chosen as the tissue reference, in likely being less affected by the inversion pulse than gray matter or CSF.

RESULTS

Figure 1 shows the seven parametric maps from a slice above the ventricles obtained with standard DKI from a representative subject. The parametric maps from the same subject using FLAIR-DKI are displayed in Figure 2, and difference images are displayed in Figure 3. The strong suppression of the CSF in the sulci is most apparent in the diffusivity maps (MD, D_∥, D_⊥).

1.

Parametric maps for the diffusion metrics of MD, D_∥, D_⊥, FA, MK, K_∥, and K_⊥ obtained from a representative subject imaged using standard DKI. The scale bar range [0, 1.0] applies to the FA map, and the range [0, 2.0] applies to all other parameters. The scale bar corresponds to units of μm²/ms for the diffusivities and is dimensionless for the FA and kurtosis measures.

2.

Parametric maps for the diffusion metrics of MD, D_∥, D_⊥, FA, MK, K_∥, and K_⊥ obtained from the same subject as in Figure 1 imaged using FLAIR-DKI. Notice that the effect of CSF suppression is most evident for the diffusivities. The scale bar conventions are as in Figure 1.

3.

Difference images for the diffusion metrics of MD, D_∥, D_⊥, FA, MK, K_∥, and K_⊥ obtained from the same subject as in Figure 1. The scale bar conventions are as in Figure 1.

Figure 4 shows examples of white matter and gray matter masks, superimposed on a b = 0 image from a representative subject. Two different masks were used in order to evaluate the sensitivity of our results from the gray/white matter segmentation procedure. As demonstrated by Figure 4, the FA mask provides a more inclusive segmentation of white matter, while the MK mask provides a more inclusive segmentation of gray matter.

4.

FA and MK masks for white matter (a, b) and gray matter (c, d). White matter was segmented using an FA threshold of 0.15 and an MK threshold of 1.0 on the standard DKI maps. For white matter, the MK mask is largely contained within the FA mask, and vice versa for gray matter. Two different masks were used to help assess the robustness of the results.

The changes in each parameter in white matter are shown in Figure 5. Use of FLAIR-DKI affects all the parameters by only a few percent. Using the MK mask, the MD, D_∥, and D_⊥ decrease by 3.2±0.3%, 2.1±0.3%, and 4.2±0.2%, respectively, with the uncertainties representing standard error estimates. In consistency with previous studies (14,15), the FA increases with application of FLAIR (by 2.4±0.1%). The kurtosis metric MK changes by 0.04±0.3%, and K_∥ and K_⊥ are found to change by 0.1±0.3%, and 1.1±0.7%, respectively. With FA segmentation, similar results are obtained. The MD, D_∥, and D_⊥ decrease by 6.6±0.5%, 4.6±0.4%, and 8.4±0.7% and the FA, MK, K_∥, and K_⊥ increase by 4.4±0.4%, 1.4±0.9%, 2.5±0.4%, and 3.7±0.9%. The use of the CSF masks to remove the CSF dominated voxels from the cortical areas does not greatly change the results. For example, after applying the CSF mask to the MK-segmented maps, the decrease in MD is still 3.2±0.3% and the change in MK is still 0.04±0.3%. With the FA-segmented maps, the MD decrease becomes 5.2±0.6%, and the MK decreases by 4.5±0.8%.

5.

Comparison of diffusion metrics measured in white matter for standard DKI and FLAIR-DKI. The mean and standard error of the mean are shown for each metric. The results are shown for both the FA and MK masks. The percent changes were calculated according to Eq. [1].

The corresponding graphs for cortical gray matter are presented in Figure 6. The differences between parameters obtained using standard and FLAIR-DKI are much greater in gray matter than in white matter, indicating a substantial partial volume effect for gray matter. With the MK mask, the MD, D_∥, and D_⊥ decrease by 34±3%, 29±3%, and 41±4%, respectively. The FA increases by 23±2%, and the MK, K_∥, and K_⊥ increase by just 6.6±2%, 8.8±2%, and 7.6±2%. Similar results are again obtained using the FA mask; the MD, D_∥, and D_⊥ decrease by 44±2%, 46±2%, and 52±3%, while the FA, MK, K_∥, and K_⊥ are found to increase by 19±2%, 7.9±3%, 11±2%, and 6.8±3%. When the CSF masks are applied to the MK-segmented images, the MD decreases by 27±4%, and the MK increases by 6.7±1.6%. For the FA-segmented images, application of the CSF masks leads to an MD decrease of 38±5% and an MK increase of 10±2%.

6.

Comparison of diffusion metrics measured in cortical gray matter for standard DKI and FLAIR-DKI. The mean and standard error of the mean are shown for each metric. The results are shown for both the FA and MK masks. The percent changes were calculated according to Eq. [1].

The SNR for the standard acquisition was 21.3±2.1, while the SNR for the FLAIR acquisition was 22.2±2.6. This close correspondence suggests that, for the FLAIR sequence, the SNR decrease caused by incomplete recovery of the longitudinal magnetization following the inversion pulse was approximately compensated by the longer TR.

DISCUSSION

Several other investigations have examined the CSF partial volume effect on estimates for diffusivity and anisotropy measures in healthy volunteers (10–17). These studies have found significant changes in FA and MD after CSF contamination is suppressed with the application of FLAIR. In our study, we extended the investigation of CSF contamination by also evaluating the D_∥, D_⊥, MK, K_∥, and K_⊥ obtained through the application of both standard DKI and FLAIR-DKI to a set of four healthy volunteers. For the MD and FA changes, our findings are qualitatively in agreement with earlier reports, with quantitative differences likely due to ROI definition and choice of sequence parameters. In particular, for the FA mask, we find an increase of 19% in cortical gray matter as compared to the increase of 88% (frontal lobe) for the study of Ma et al. (14) and 7% for the study of Bhagat et al. (15). For our study, we intentionally considered an ROI that included the CSF within the sulci, since CSF is often difficult to reliably segment from cortical gray matter in diffusion-weighted images. As demonstrated by Latour and Warach (13), the effect of CSF contamination on diffusion metrics depends sensitively on the parenchyma volume fraction. We note that the MD values measured with DKI tend to be higher than those measured with DTI, due to the higher order modeling of the signal decay (34).

A key result of this study is the smaller CSF contamination effect observed for the kurtosis measures, as compared with conventional DTI metrics. This is true in white matter, but more dramatically so for cortical gray matter. For gray matter, the use of FLAIR led to diffusivity reductions of 29 to 52%, but kurtosis increases of only 7 to 11%. The kurtosis parameters may then be relatively specific indicators of true alterations in tissue parenchyma microstructure due to pathology, since they are less affected by changes in CSF partial volume caused, for example, by loss of tissue. Of course for any particular application, the comparative utilities for assessing tissue alterations of diffusivities and kurtoses, which are after all distinct physical quantities, depend also on the effect sizes of these parameters.

One reason the differences are small in white matter may be due to the fact that white matter is largely surrounded by gray matter and thus not as directly influenced by CSF partial voluming. Nonetheless, the percent changes in MK, K_∥, and K_⊥ between standard and FLAIR-DKI are smaller than the changes in the other parameters. This observation is most apparent when using the more conservative MK mask. The larger percent changes observed in gray matter are most likely due to the proximity of CSF in the sulci. Nonetheless, even when CSF dominated voxels are explicitly eliminated, the relative consistency of the kurtosis parameters is still apparent.

A potential explanation for the robustness of the kurtosis metrics can be put forth in terms of an elementary multiple compartment model. Suppose the contents of a particular voxel consist of N non-exchanging, Gaussian compartments, one of which corresponds to CSF. The diffusivity, along a particular direction, is then given by

$$D=\sum _{i=1}^N{f}_i{D}_i,$$ [2]

where D_i is the diffusivity for the ith compartment and f_i is the corresponding water fraction. Typically, the CSF compartment would have a substantially larger D_i than the parenchymal compartments, since water diffuses much more rapidly in CSF than inside brain tissue. Thus, suppressing the CSF compartment with FLAIR may result in a large decrease in D even if the CSF represents only a small part of the voxel. For this same model, the kurtosis is given by (18,20)

$$K=3\frac{{\delta }^{2}D}{D^2},$$ [3]

with δ²D being the variance of the compartmental diffusivities expressed explicitly by

$${\delta }^{2}D=\sum _{i=1}^N{f}_i{(D_i-D)}^2.$$ [4]

When the CSF is suppressed, it is then likely that both D² and δ²D decrease so that the relative change in their ratio, which is just K/3, may be less than for the individual quantities. Another conceivable reason for the smaller observed changes in kurtosis is that there may be a fortuitous effect from systematic errors in the kurtosis estimation, which can in practice depend somewhat on a voxel's composition (20).

A potential confounding factor in comparing standard and FLAIR DKI are differences in SNR, which could result in discrepant systematic deviations for the parameter estimates (19). However, this is greatly mitigated with our protocol, as the SNR was very similar for the two acquisitions. Another possible confounding effect is that the inversion pulse influences white and gray matter differently due to their distinct T1 values, but this is only a concern in voxels with significant amounts of both types of tissue.

That CSF partial voluming can be substantially suppressed with the application of FLAIR has been demonstrated here, as in prior studies (10–17). However, this is not a feasible technique for most applications due to long acquisition times (44), although a more time efficient FLAIR diffusion method has been recently suggested (45). Therefore, the low sensitivity of the diffusional kurtosis metrics to CSF contamination may be an important advantage in clinical studies utilizing standard diffusion-weighted sequences. For DTI, bi-tensor analysis of standard DTI data has been proposed as an alternative to FLAIR for separating tissue from free water diffusion (9,44). It would be of interest to investigate whether this method could also be extended to DKI.

In conclusion, this work demonstrates that measured diffusional kurtosis metrics are relatively robust with respect to CSF contamination, which helps to support the reliability of standard DKI in brain. When CSF is suppressed using FLAIR-DKI, smaller changes are observed in the kurtosis metrics of MK, K_∥ and K_⊥ than in the conventional DTI metrics of MD, D_∥, D_⊥ and FA. While the observed changes are in all cases less than 10% in white matter, in cortical gray matter the diffusivity indices change by 19 to 52% compared to 7 to 11% for the kurtosis metrics. These findings suggest that the kurtosis may be a more specific indicator of alterations associated with brain tissue microstructure (e.g., neurodegeneration), provided the kurtosis parameters have comparable effect sizes. Analysis of these specific parameters using DKI can thus potentially improve the sensitivity of diffusion-weighted imaging to tissue alteration caused by neuropathology.