Diffusional Kurtosis Imaging in the Diffusion Imaging in Python Project

Abstract

Diffusion-weighted magnetic resonance imaging (dMRI) measurements and models provide information about brain connectivity and are sensitive to the physical properties of tissue microstructure. Diffusional Kurtosis Imaging (DKI) quantifies the degree of non-Gaussian diffusion in biological tissue from dMRI. These estimates are of interest because they were shown to be more sensitive to microstructural alterations in health and diseases than measures based on the total anisotropy of diffusion which are highly confounded by tissue dispersion and fiber crossings. In this work, we implemented DKI in the Diffusion in Python (DIPY) project - a large collaborative open-source project which aims to provide well-tested, well-documented and comprehensive implementation of different dMRI techniques. We demonstrate the functionality of our methods in numerical simulations with known ground truth parameters and in openly available datasets. A particular strength of our DKI implementations is that it pursues several extensions of the model that connect it explicitly with microstructural models and the reconstruction of 3D white matter fiber bundles (tractography). For instance, our implementations include DKI-based microstructural models that allow the estimation of biophysical parameters, such as axonal water fraction. Moreover, we illustrate how DKI provides more general characterization of non-Gaussian diffusion compatible with complex white matter fiber architectures and grey matter, and we include a novel mean kurtosis index that is invariant to the confounding effects due to tissue dispersion. In summary, DKI in DIPY provides a well-tested, well-documented and comprehensive reference implementation for DKI. It provides a platform for wider use of DKI in research on brain disorders and cognitive neuroscience research. It will ease the translation of DKI advantages into clinical applications.

Diffusional Kurtosis Imaging in the Diffusion Imaging in Python Project

“…Moreover, our simulations ( Figure 3 ) shows that DKI parameters can not only suffer from low precision but also biased by Rician noise even at typically diffusion data SNRs ( 20-40). As illustrated in Figure 5 , thermal noise biases can manifest as implausible negative estimates “black voxels” in standard DKI maps, particularly in regions where diffusivities are low (Tabesh et al, 2011 ; Henriques, 2012 ; Kuder et al, 2012 ; Veraart et al, 2013 ). Implausible negative kurtosis estimates can also originate from effects of different image artefact such as Gibbs Ringing artefacts as explained by Perrone et al ( 2015 ) and Veraart et al ( 2016a ).…”

An open, analysis-ready, and quality controlled resource for pediatric brain white-matter research

“…These processes create bundle profiles, in which diffusion measures are quantified and averaged along twenty-four major fiber tracts. We retain only the mean diffusivity (MD) and the fractional anisotropy (FA) from a diffusion kurtosis imaging (DKI) model ( Jensen et al, 2005 ), implemented in DIPY ( Henriques et al, 2021 ), and impute missing bundles using median imputation as implemented by scikit-learn ’s . Because the HBN-POD2 bundle profiles exhibit strong site effects ( Richie-Halford et al, 2021 ), we used the ComBat harmonization method to robustly adjust for site effects in the tract profiles.…”

“…, implemented in DIPY(Henriques et al, 2021), and impute missing bundles using median imputation as implemented by scikit-learn's SimpleImputer class. Because the HBN-POD2 bundle profiles exhibit strong site effects (Richie-Halford et al, 2021), we used the ComBat harmonization method to robustly adjust for site effects in the tract profiles.…”

Improved reproducibility of diffusion kurtosis imaging using regularized non-linear optimization informed by artificial neural networks

“…Initial positions were computed by solving the problem with ordinary least squares. The result of the fit was considered successful if AKC(n) ≥ 0 for all n in a 45-point 8-design (Hardin and Sloane, 1996;Henriques et al, 2021a).…”

“…where ÂK, MK, and RK are the predicted kurtosis values, α is a constant controlling the magnitude of the regularization terms, and m, a, and r are functions for numerically computing mean, axial, and radial kurtosis, respectively (Henriques et al, 2021a). Initial positions corresponded to axially symmetric diffusion and kurtosis tensors (Hansen et al, 2016) aligned with the principal diffusion direction and with plausible magnitudes computed from the results of the steps described in sections 3.1.1 and 3.1.2.…”