As a basis for DTI data analysis, it has to be noted that WM tracts in the central nervous system consist of densely packed axons in addition to various types of neuroglia and other small populations of cells (Nieuwenhuys et al., 2019). The axonal membrane as well as the well-aligned protein fibers within an axon restrict water diffusion perpendicular to the fiber orientation, leading to anisotropic water diffusion in brain WM (Moseley et al., 1990). Myelin sheaths around the axons may also contribute to the anisotropy for both intra-and extracellular water (Mori and van Zijl, 2002). Diffusion anisotropy is mainly caused by the orientation of fiber tracts in WM and is influenced by its micro-and macrostructural features. Of the microstructural features (Basser et al., 1994; Mattiello et al., 1994), intraaxonal organization appears to be of greatest influence on diffusion anisotropy, besides the density of fiber and cell packing, degree of myelination, and individual fiber diameter. On a macroscopic scale, the variability in the orientation of all WM tracts in an imaging voxel influences its degree of anisotropy (Pierpaoli and Basser, 1996). DTI provides two types of information about the property of water diffusion: first, the orientation-independent extent of diffusion anisotropy (Pierpaoli and Basser, 1996) and second, the predominant direction of water diffusion in image voxels, i.e., the diffusion orientation (Pajevic and Pierpaoli, 1992). Since there are several challenges in displaying tensor data, the concept of diffusion ellipsoids has been proposed (Basser et al., 1994). The Eigendiffusivities of these ellipsoids represent the unidimensional diffusion coefficients in the main direction of diffusivities of the medium, i.e., the main axis of the ellipsoid represents the main diffusion direction in the voxel which coincides with the direction of the fibers, while the eccentricity of the ellipsoid provides information about the degree of anisotropy and its symmetry. Therefore, diffusion anisotropy metrics such as the fractional anisotropy (FA) could be defined for the parameterization of the voxel tensors (Le Bihan et al., 2001). Advanced techniques like Q-ball imaging (Tuch, 2004) may provide more sensitive white matter descriptors in single patients and lead to generalized measures of fractional anisotropy (Corbo et al., 2014). Thus, the clinical role of DTI in various disease processes, especially NDDs, is emerging (Tae et al., 2018).

These DTI-based metrics have shown to map age-related alterations over the life span in the human brain; such age-dependent changes also seem to exhibit regional differences with respect to the brain anatomy (Salat et al., 2005; Westlye et al., 2010). Furthermore, a regional tract-specific age dependency which requires also non-linear corrections for different age ranges has been demonstrated (Behler et al., 2021). Based on these findings, DTI also acts as a cross-validation technique for age-dependent cerebral blood flow alterations in the human brain (Damestani et al., 2024).

Standardized DTI protocols will, at least currently, be performed at 1.5 T or 3.0 T clinical scanners, because scanners with higher magnetic field (ultrahigh field scanners) would provide better signal-to-noise ratio but are still rare and hardly available for multicenter trials. The protocols include full-brain coverage using a 2-D echo planar imaging sequence; about axial 70 slices, slice thickness 2.0 mm, voxel size 2.0 × 2.0 × 2.0 mm³, field of view 256 × 256 mm², matrix 128 × 128, five b₀ images, and 30 or more diffusion gradient directions with a b value of 1,000 s/mm². The recording of more than one b₀ image could help to improve the signal-to-noise-ratio (SNR). Good examples of standardized prospective multicenter protocols are provided in (Hobbs et al., 2015 – HD; Nir et al., 2015 – AD; Kalra et al., 2020 – ALS). However, a major drawback of DTI is that it assumes a single water pool with Gaussian diffusion in each voxel and does not account for structural heterogeneity which reduces the specificity of the derived indices (Alexander et al., 2001; Basser and Jones, 2002), neurite orientation dispersion and density imaging (NODDI) is a multishell diffusion technique (Zhang et al., 2012) that assumes further types of microstructural environments and could easily be added to a standardized DTI protocol, if b-values of 2000 s/mm² are available.

For the application of fiber tracking (Mori and van Zijl, 2002), differently oriented fiber bundles inside one voxel are incorrectly modeled by a single tensor. High Angular Resolution Diffusion Imaging (HARDI) aims at using more complex models, such as a two-tensor model, for estimating two fiber bundles (Caan et al., 2007). Although standard DTI is an established default tool, acquisition with stronger diffusion weightings beyond the DTI regimen is now possible by NODDI, that way enabling even more detailed characterization of tissue microstructures, e.g., in the early diagnosis of AD (Takahashi et al., 2024) or tau deposition in AD (Weston et al., 2023).

The schedule of visits could be optimized taking possible costs and patient burden into account. Repeated DTI scans during one given subject’s visit could significantly improve the effect size (Behler et al., 2022a). Thus, it is recommended to increase the number of DTI scan repetitions at one visit. Note that the scans should be independent, i.e., subject repositioning in the scanner is recommended.

For longitudinal studies in NDD with a rapidly progressive disease course, e.g., ALS, one baseline and two follow-up visits were advised over a total observation period of about 12 months (time interval between baseline and follow-up 2) to monitor the course of disease (e.g., Cardenas-Blanco et al., 2016; Kassubek et al., 2018; Kalra et al., 2020). For longitudinal studies in NDD with a slowly progressive disease course, multiple follow-up scans are possible and could even be spread over several years, e.g., in PD with one baseline and two follow-up visits 6 years apart (Shih et al., 2023).

Independent of the progression rate, it has to be noted that, if one baseline and two follow-up scans are recorded, the first follow-up visit is suggested to balance observing change and minimizing attrition so that the individual schedule should be such that the timing of follow-up 1 does not bisect the entire observation period between baseline and follow-up two. The unfavorable choice of a symmetrical bisectioning of the observation interval was demonstrated in a cohort of HD patients with one baseline and 2 follow-up scans (Müller et al., 2021). The shortcoming with splitting the observation time into two identical periods is that regression models, due to their intrinsic properties, underweight follow-up visit 1, with the result that the progression slope is mainly determined by baseline and follow-up 2.

In addition to common MRI artifacts (truncation, aliasing, chemical shift, banding, pile-up, blurring, spikes, etc.) (Krupa and Bekiesińska-Figatowska, 2015), there are specific challenges that one may encounter when using MRI scanner gradient hardware for diffusion MRI, especially in terms of eddy currents and sensitivity to motion (Le Bihan et al., 2006). According to an established quality control protocol (Dubois et al., 2014; Müller et al., 2014), corrupted gradient directions as well as motion artifacts could be excluded and corrupted slices or volumes could be resampled from further analysis prior to correction of eddy current-induced geometric distortions (Shen et al., 2004). In particular, if the analysis focuses on one parameter like FA, the omission of single GD has only little effect on the result and thus the omission of corrupt GD allows a more precise determination of the resulting values.

Considering the effect of experimental noise on DTI metrics (Bastin et al., 1998; Seo et al., 2019), statistical analysis is provided by several software packages [e.g., Statistical Parametric Mapping (SPM) (Penny et al., 2006), FMRIB Software Library (FSL) (Jenkinson et al., 2012), Tensor Imaging and Fiber Tracking (TIFT) (Müller et al., 2007)]. Analysis could be performed as unbiased whole brain-based voxelwise comparison of DTI metrics (Müller et al., 2014), or as hypothesis-guided region of interest (ROI) or tract of interest (TOI) analyses [tract-based spatial statistics (TBSS) (Smith et al., 2006) or tractwise fractional anisotropy statistics (TFAS) (Mueller et al., 2007)]. Analyses pipelines combining several toolboxes have also been introduced (e.g., Cui et al., 2013; Soares et al., 2013).

The principles of microstructural analysis by DTI have been taken use of in many mono-and multicenter studies in NDD:

The cross-sectional comparison of NDD patient groups to healthy controls aims at the definition of specific brain regions and the related patterns of alterations. Beyond this initial cross-sectional research in specific NDDs, a “pseudo-longitudinal” analysis of DTI could be performed: cross-sectional results could also be further analyzed by taking the disease status of the individuals as time-axis in order to map increasing alterations or spreading patterns during the course of ALS (Kassubek et al., 2014; Müller et al., 2023) or association patterns similar to the neuropathological Braak staging of AD (Wen et al., 2021) could be analyzed. Meta-analyses of DTI metrics in PD contribute to increasing the knowledge of PD pathophysiology by addressing the possibility of follow-up of the disease severity and associated brain structural modulations using in vivo imaging (Atkinson-Clement et al., 2017).

In case of unequal sample sizes, the Welch-test should be used which is an adaption of Student’s t-test and is more reliable when the two samples have unequal variances. Welch’s t-test defines the statistic by

with σ N D D , controls 2 = σ N D D 2 / N N D D + σ controls 2 / N controls

Prior to statistical analysis, input data should be tested for normal distribution. In case of not normally distributed data, the samples should first be checked with regard to their distribution and the contributing data (see subsection “Selection of subject samples for clinical trials”) and, if not normally distributed, non-parametric testing, e.g., Mann–Whitney U-test, should be applied.

Previous studies that analyzed longitudinal alterations in ALS patients (Zhang et al., 2016; Kassubek et al., 2018; Kalra et al., 2020; Shih et al., 2023) calculated the effect size of longitudinal differences in a given DTI metric by

with

Δ σ group 2 = 1 N ∑ i N S i , F U P − S i , B L group − Δ μ group 2

Δ μ N D D , Δ μ controls are the average longitudinal alterations in controls and in NDD patients at the group level. Note: in this approach a longitudinal time interval normalization is necessary (Kassubek et al., 2018).

Eq. 13 can be written as

Now, the effect size of a two sample (NDD patients and controls) study is

That way the noise and inter-individual variability are contributing 2 N controls + 2 N N D D times .

This approach also takes longitudinal alterations in controls into account, that way, mainly incorporating additional noise components due the information input of longitudinal scans of controls. Since the target is the detection of alterations in NDD patients (see subsection “What can we gain from longitudinal scans of controls?”), the alterations of inter-subject variability of controls Δ F i (as obtained from longitudinal scans of controls) is off target and would only lead to an additional noise component in the results.

An alternative approach (e.g., Cardenas-Blanco et al., 2016; Loane et al., 2016) is to analyze data without follow-up control visits; then, Eq. 15 simplifies to

Then the noise and inter-individual variability are contained 2 N N D D times. This approach (ideally with a large control sample at baseline) shows a higher impact on the effect size as Δ σ controls 2 = 0 .Thus

However, this approach without follow-up of the controls is particularly suitable for cross-sectional comparisons of each visit to a (large) database of controls (Eq. 7) so that optimal effect size can be obtained at any visit. This is particularly suitable for a 3-D visualization of longitudinal alterations (unbiased whole brain-based voxelwise analysis) or for the display of longitudinal changes in defined structures (e.g., Cardenas-Blanco et al., 2016). That means that higher effect sizes could be gained when data of each visit (of NDD patients) are compared to a (large and optimized) control sample at baseline.

Multiparametric or multimodal imaging has a high impact on the quality and reliability of results. The investigation of the relationship between WM alterations with further factors (atrophy, vascular disease, or, e.g., amyloid burden) and clinical features increases the set of analyzed parameters and thus offers possibilities to apply methods of multiparametric analyses, e.g., in AD and Lewy body dementia (Donaghy et al., 2020). Recent multiparametric analyses showed no association between multiple biomarkers of cerebral microvascular function and WM connectivity (using DTI to quantify the number and organization of WM connections) in a large cohort of some thousand participants (Beran et al., 2024), however, further MRI techniques to assess blood–brain-barrier permeability or cerebrovascular reactivity and microvascular perfusion at the tissue level could be combined with WM connectivity analysis of DTI (Beran et al., 2024). Multimodal association analyses of DTI with amyloid beta and tau positron emission tomography (PET) show potential to provide information relating to underlying tau deposition in AD (Weston et al., 2023).

DTI can be used to tract-wise map correlates of the sequential disease progression and, therefore, to assess disease stages, e.g., in ALS (Kassubek et al., 2014, 2018). A technical improvement in reliability of the analysis was reached by applying a multistage classifier based on Bayesian statistics (Behler et al., 2022b) with the significant advantage of Bayesian statistics for multimodal issues by incorporating prior knowledge about the patient into the algorithm.

A multiparametric set of variables (including the predictive value of microstructural integrity) could be used for association with clinical phenotype features like cognitive performance (Power et al., 2019), the improvement of diagnostic accuracy (Piersson et al., 2021), or to assess the predictive value on survival (Agosta et al., 2019; Kuan et al., 2023) in NDD.

The predictive value of microstructural integrity has been used by artificial intelligence (AI) as machine learning (ML) tools; especially convolutional neural networks (CNN), support vector machines (SVM), and random Forrest models (RFM) successfully contributed to neuroimaging studies in NDDs.

Especially in multicenter trials, different factors may contribute to the variability of DTI data of controls and NDD patients. Although the precise influence of each source of variation could not be delineated, investigating group FA differences between patients and controls on systematic between-center differences should be investigated prior to pooling across centers. Furthermore, center-specific sources of variability on DTI metrics, e.g., scanner-specific variability, environmental noise and specific factors such as scanning time, might be present in single center studies but will only slightly influence comparisons at the group level (Müller et al., 2013). Although multicenter studies could allow a direct merging of DTI metrics across centers (Müller et al., 2013), especially in case of different scan protocols, merging of DTI metrics requires the application of a strategy to regress out confounders such as field strength, echo time, or number of gradient directions (Müller et al., 2016). Especially, FA is influenced by the voxelsize (Oouchi et al., 2007) and thus needs harmonization in case of different scan protocols in different centers. Harmonization could be performed by 3-D voxelwise linear correction matrices (Rosskopf et al., 2015); recently, a post-processing technique based on rotation invariant spherical harmonics features was introduced to mitigate cross-scanner differences in DTI metrics (De Luca et al., 2022).

The development of a Domain Shift Analyzer for MRI (DSMRI) which was designed explicitly for multicenter MRI datasets (Kushol et al., 2023a,b) allows for the identification of NDD as demonstrated for ALS cases (Kushol et al., 2023c), thus offering the possibility to investigate neuroimaging modalities like functional MRI (fMRI) and DTI within a similar environment.

In addition to the hardware and software already tailored to optimize the SNR of individual recordings by the manufacturer, significant improvements can be achieved in the selection of study participants, the design of the studies, the application of the DTI protocols and the data analysis in pre-and postprocessing with regard to the effect size and thus the sample size in longitudinal clinical studies/trials (Figure 2).

This summary of recommendations was conducted to set up a framework for a standardized DTI scanning and DTI analysis protocol in studies in NDD (Figure 2):

For neuroimaging studies as a part of clinical therapeutic trials in NDD, general clinical recommendations should be considered: first, to obtain a homogeneous study population, i.e., the disease progression rate and also the genetic phenotyping of NDD patients should be part of the inclusion/exclusion criteria. Second, primary endpoint and treatment duration must be matched to the study population, especially for slowly progressive NDD patients.

The potential of DTI-based metrics as a non-invasive progression marker during the disease progression in NDD is influenced by sample size, scheduling of baseline and follow-up sessions, and measurement uncertainty on the statistical power. Especially for measurements with a limited SNR, for example, due to subject-related factors (Müller et al., 2013), the application of the recommendations of this review will potentially strengthen the reliability of the FA values, in line with SNR improvement by signal-averaging during individual scans (Farrell et al., 2007; Seo et al., 2019). Vice versa, the increased statistical power of a DTI protocol means that lower sample sizes might suffice to measure small effects and/or effects after a short time, respectively (Behler et al., 2022a). Another critical factor is the composition of the groups and the selection of the scans to be included in the analysis. Prior to analysis, the restriction to carefully selected patients and controls saves costs and could increase effect sizes, that way also reducing patient burden and increasing subject motivation. It has to be added that data collection (selection of participants) as well as data quality have to be optimized during the process of data acquisition because subsequent ex post facto correction for age, different study protocols and quality control of movements and other artifacts can only partially correct for such deviations and could, thus, never reach the data quality of optimized (standardized) recordings and subject selections.

The technical aspects of planning medical trials listed here can generally be implemented with little effort, but when implemented in full or in part, they have the potential for a significant improvement in results with a simultaneous potential reduction in costs.

Based on characteristic alteration patterns of cross-sectional comparison of NDD patients to controls, DTI appears to be useful at distinguishing frontotemporal dementia (FTD) from patients with AD, also at the individual level (Grossman, 2010). DTI has shown to be able to classify subjects diagnosed with PD, atypical parkinsonism, and essential tremor and to distinguish them from control subjects (Prodoehl et al., 2013). DTI indicators of white matter impairment have the potential to emerge as useful clinical tools for differentiating diagnostic groups in studies of AD, MCI, and normal aging (Nir et al., 2013, 2015). ALS and FTD encompass a clinical, pathological and genetic continuum, and ALS could be mainly distinguished from ALS with frontotemporal dementia (ALS-FTD) and behavioral variant FTD by different white matter microstructure alteration pattern, especially corticospinal tract degeneration (Lillo et al., 2012).

DTI has been used to study the effects of NDD on neural pathways which may lead to more reliable and early diagnosis of these diseases as well as a better understanding of how they affect the brain. In AD, ML methods were applied for defining DTI metrics (Konukoglu et al., 2016; Lombardi et al., 2020; Xu et al., 2021; Agostinho et al., 2022) to characterize MCI (Velazquez and Lee, 2022; Zhou et al., 2022a,b; Cheng et al., 2023) and to predict AD early (Savarraj et al., 2022). The characterization of MCI and cognitive impairment in PD (Xu et al., 2021; Yang Y. et al., 2022; Chen B. et al., 2023; Huang et al., 2023) or the investigation of progression in PD (Prasuhn et al., 2020; Yang et al., 2021a,b) has also been addressed by the application of ML methods. Furthermore, ML was applied to the differentiation of parkinsonian syndromes (Haller et al., 2012; Du et al., 2017; Chougar et al., 2021; Talai et al., 2021). Since first applications of ML methods to ALS (Welsh et al., 2013; Sarica et al., 2017), ML was used to improve diagnostic accuracy (Kocar et al., 2021; Behler et al., 2023) and clinical associations (Li et al., 2021). For the evaluation of imaging biomarkers in HD, ML methods have been applied for many years (Klöppel et al., 2008; Rizk-Jackson et al., 2011; Hu B. et al., 2023). Thus, ML and AI are exponentially improving medical imaging and diagnosis: by ML techniques (especially SVM or RFM), the combination of multiparametric/multimodal imaging data allow for a multidimensional analysis beyond (simple) association analysis (Tae et al., 2018; Weston et al., 2023; Takahashi et al., 2024). Moreover, CNNs might enable estimations of the risk status, and thus, prediction of the disease course including survival in NDDs (Agosta et al., 2019; Kuan et al., 2023). Altogether, it seems safe to expect that ML methods will help to define the future of DTI-based analysis of WM integrity in the brain and its affectations by NDD.

Technical considerations are presented that improve the information contents in DTI recordings and subsequent statistical analysis in clinical DTI-based neuroimaging trials in NDD. In summary, this study and the included recommendations potentially might enhance the role of DTI as a biomarker in NDD by standardized scan protocols and analysis cascades.

Especially AI-based techniques improved the performance of multimodal MRI including DTI in survival prediction and the prognostic value in NDD (Agosta et al., 2019). Furthermore, AI-aided DTI analysis provides high-precision automatic diagnosis and simultaneously output feature probability maps to provide clinical auxiliary diagnosis in NDD (Deng et al., 2023). Thus, there is the urgent need for the translation of advanced brain MRI techniques into clinical practice, e.g., for an assessment of prognostic factors and a stratification of patients in the design of pharmacological trials (Agosta et al., 2019). That way, the predictive value of microstructural integrity by AI-based algorithms, especially in a multiparametric or multimodal setting, could be developed as a non-invasive in vivo biomarker in NDD.