Purpose: An increasing number of studies have aimed to compare diffusion tensor imaging (DTI)-related parameters [e.g., mean diffusivity (MD), fractional anisotropy (FA), radial diffusivity (RD), and axial diffusivity (AD)] to complementary new indexes [e.g., mean kurtosis (MK)/radial kurtosis (RK)/axial kurtosis (AK)] derived through diffusion kurtosis imaging (DKI) in terms of their discriminative potential about tissue disease-related microstructural alterations. Given that the DTI and DKI models provide conceptually and quantitatively different estimates of the diffusion tensor, which can also depend on fitting routine, the aim of this study was to investigate model- and algorithm-dependent differences in MD/FA/RD/AD and anisotropy mode (MO) estimates in diffusion-weighted imaging of human brain white matter.
Methods: The authors employed (a) data collected from 33 healthy subjects (20-59 yr, F: 15, M: 18) within the Human Connectome Project (HCP) on a customized 3 T scanner, and (b) data from 34 healthy subjects (26-61 yr, F: 5, M: 29) acquired on a clinical 3 T scanner. The DTI model was fitted to b-value =0 and b-value =1000 s/mm(2) data while the DKI model was fitted to data comprising b-value =0, 1000 and 3000/2500 s/mm(2) [for dataset (a)/(b), respectively] through nonlinear and weighted linear least squares algorithms. In addition to MK/RK/AK maps, MD/FA/MO/RD/AD maps were estimated from both models and both algorithms. Using tract-based spatial statistics, the authors tested the null hypothesis of zero difference between the two MD/FA/MO/RD/AD estimates in brain white matter for both datasets and both algorithms.
Results: DKI-derived MD/FA/RD/AD and MO estimates were significantly higher and lower, respectively, than corresponding DTI-derived estimates. All voxelwise differences extended over most of the white matter skeleton. Fractional differences between the two estimates [(DKI - DTI)/DTI] of most invariants were seen to vary with the invariant value itself as well as with MK/RK/AK values, indicating substantial anatomical variability of these discrepancies. In the HCP dataset, the median voxelwise percentage differences across the whole white matter skeleton were (nonlinear least squares algorithm) 14.5% (8.2%-23.1%) for MD, 4.3% (1.4%-17.3%) for FA, -5.2% (-48.7% to -0.8%) for MO, 12.5% (6.4%-21.2%) for RD, and 16.1% (9.9%-25.6%) for AD (all ranges computed as 0.01 and 0.99 quantiles). All differences/trends were consistent between the discovery (HCP) and replication (local) datasets and between estimation algorithms. However, the relationships between such trends, estimated diffusion tensor invariants, and kurtosis estimates were impacted by the choice of fitting routine.
Conclusions: Model-dependent differences in the estimation of conventional indexes of MD/FA/MO/RD/AD can be well beyond commonly seen disease-related alterations. While estimating diffusion tensor-derived indexes using the DKI model may be advantageous in terms of mitigating b-value dependence of diffusivity estimates, such estimates should not be referred to as conventional DTI-derived indexes in order to avoid confusion in interpretation as well as multicenter comparisons. In order to assess the potential and advantages of DKI with respect to DTI as well as to standardize diffusion-weighted imaging methods between centers, both conventional DTI-derived indexes and diffusion tensor invariants derived by fitting the non-Gaussian DKI model should be separately estimated and analyzed using the same combination of fitting routines.