Multiple sclerosis (MS) is a severe central nervous system disease impacting > 2.8 million people worldwide (Msif Tmsif., 2020). Focal lesions are the hallmark of MS pathology, characterized by several changes, including demyelination, axonal injury, and inflammation (Reich et al., 2018). In addition, MS pathology alters tissue integrity not only in the normal-appearing tissue, but also within the territory of established lesions, leading to heterogeneous damage as seen in chronic active lesions. Ongoing tissue damage in the latter is believed to play a major role in the relentless progression of patient disability in MS (Reich et al., 2018). However, robust measurements of the complex types of MS pathology are still missing. Diffusion magnetic resonance imaging (dMRI) serves as a promising tool for in vivo assessment of tissue microstructure (Inglese and Bester, 2010). Moreover, with advances in dMRI techniques, multi-dimensional analysis of tissue properties becomes possible, including localized analysis of lesion activities concerning specific white matter tracts (Inglese and Bester, 2010).

Diffusion tensor imaging (DTI) is a common dMRI method that reflects the orientational dependence of diffusion associated with individual image voxels (Tournier, 2019). Diffusion tensor imaging uses a tensor model, which defines diffusion anisotropy according to three perpendicular diffusion axes originating from a voxel (Tournier, 2019). Several studies have shown that the mean diffusivity (MD) and fractional anisotropy (FA) of DTI detect demyelination and axonal loss in MS, which differentiates lesions from the normal-appearing white matter (NAWM) (Schmierer et al., 2007). Within the lesion context, there is evidence showing increased MD in the lesion core versus perilesional white matter (Klistorner et al., 2018). However, there are no systemic studies of core versus periphery of lesions in MS, and use of a single tensor model in DTI, which does not provide compartmental information, limits its ability to detect specific processes of neuronal pathology. High angular resolution diffusion imaging (HARDI) tackles this problem by using advanced models which enable intra-voxel analysis through increased angular sampling of diffusion signals acquired typically with different diffusion weightings (multi-shell HARDI) (Descoteaux, 2015). A range of multi-shell HARDI models exist, divided mainly by the approach used to model the properties of nerve fibers, such as the cylinder model used in ActiveAx (Alexander et al., 2010), and the stick model in neurite orientation dispersion and density index (NODDI) (Zhang et al., 2012). But acquiring multi-shell HARDI data is not always practical due to time constraints, particularly in a clinical setting.

An alternative approach to multi-shell HARDI is modeling densely sampled data acquired using one diffusion weighting, namely, single-shell HARDI (ssHARDI) (Taquet et al., 2013). In general, dMRI acquired with gradient sampling orientations ≥ 45 is sufficient for this approach (Descoteaux, 2015). Current investigations of ssHARDI mainly focus on its neurite orientation characterization potential (Taquet et al., 2013). While some work has explored the possibility of generating additional compartmental metrics (Grussu et al., 2014; Magnollay et al., 2014), the ability of ssHARDI for advanced intra-voxel analysis of tissue properties remains unclear, particularly based on typical DTI data.

In addition to advances in diffusion modelling, there have also been considerable improvements in tractography investigations based on dMRI (Tournier, 2019). Traditionally, individual white matter tracts (e.g., corticospinal tract) traced by streamlines of tractography form the mainstay for localized analysis of diffusion metrics. However, the local orientation information represented directly by the streamlines may be also critical indicators of tissue properties, including streamline counts and streamline termination frequency, providing a new dimension of microstructural measurements (Ukmar et al., 2012).

The purpose of this study was to test the feasibility of ssHARDI using clinically available dMRI, and investigate how advanced metrics from ssHARDI and DTI tractography compare to traditional DTI measures in assessing MS pathology. The investigations used a machine learning technique, support vector machine (SVM), to evaluate feature importance through several tissue classification tasks. Specifically, based on 3D regions of interest (ROIs), there were two classifications within individual patients: (1) lesion versus the corresponding contralateral NAWM, to obtain a basic understanding of the sensitivity of the features to MS pathology; and (2) lesion core versus shell, to evaluate the detectability of intra-lesion pathology by the features. Between patients, there were also two analyses: lesion comparisons, between patients having high (top 25%) and low (bottom 25%) lesion counts as a surrogate for disease activity; and NAWM comparisons, between the same subject groups.

Using commonly available dMRI data, we showed the feasibility of conducting ssHARDI analysis and the complementary value of parameters from relatively simple and complex diffusion models for assessing MS pathology. It appears that tissue alignment and orientation measures from ssHARDI and DTI are particularly sensitive to the existence of lesions in a patient, and to subtle structural differences between NAWM areas (specifically the ODF energy metric) between patients with high versus low disease activity. In contrast, tissue diffusivity and alignment metrics from DTI are the best in distinguishing intra-lesion pathology, and in separating lesion activity between high and low disease patients, especially MD and FA.

The ability to evaluate sub-voxel-based tissue properties through ssHARDI modeling would be important in a clinical setting. However, ssHARDI studies often face the challenge of not obtaining enough diversity of measures for evaluation, unlike multi-shell models. To compensate, a prior study applied a partial, 2-compartment model and showed that both neurite density and dispersion indices from ssHARDI decreased in the NAWM of MS patients, and the reduction in the right internal capsule correlated with MS duration (Magnollay et al., 2014). In the present study, we undertook an alternative approach with the assistance of an efficient modeling method, AMICO. This provided several sub-voxel measures of neurite diameter, density, and dispersion. Both visual and quantitative analyses show that the variances at a ROI level are equivalent between these measures derived using either 1-shell or multi-shell data, and using either online or own DTI data in ssHARDI modeling. In addition, based on the distribution of diffusion at all sampling directions in a voxel, we also created a new measure of tissue organization, ODF energy, to increase the capacity of ssHARDI.

To maximize the understanding of diffusion metrics in tissue assessment, our study considered several modeling approaches, 3 regarding ssHARDI alone: ActiveAx, NODDI, and diffusion ODF. Multi-model analysis is critical for assessing complex pathologies as observed in MS (Reich et al., 2018). However, this strategy also led to a large volume of parameters. A caveat of multiparametric approaches (Cercignani and Bouyagoub, 2018) is the difficulty of assigning importance to parameters that have overlapping sensitivities. Here we took advantage of SVM, a well-recognized machine learning method. The SVM-RFE has shown to be robust to feature redundancy and model overfitting, and in a linear form, the SVM can assign unique coefficients to each parameter, thereby identifying the specific contribution of a parameter to individual classification tasks (Guyon et al., 2002).

In lesion–NAWM analysis, nearly all top rankings were tissue orientation and alignment metrics, particularly the top 3: ODF energy, FA, and ODI. Based on definition, ODF energy measures the orientational complexity of a structure. The higher the value, the more misalignment in the structure. FA detects changes in both axonal density and alignment anisotropy (Zhang et al., 2012), and therefore is partially explainable by ODI. The leading performance of these orientation features in this analysis may indicate that the most critical changes following lesion formation in MS are tissue damage, such as inflammatory demyelination and axonal injury. A direct consequence of this pathology is increased structural heterogeneity, rather than simple alterations in neurite density. This is consistent with prior findings showing decreased FA in nearly all lesion studies in MS compared to the NAWM (Schmierer et al., 2007; Inglese and Bester, 2010), and ODI changes with lesion formation (Grussu et al., 2017; Schneider et al., 2017) and repair (Luo et al., 2019), although the consistency of ODI changes deserves further validation (Spano et al., 2018). Notably, optimal classification between lesion and NAWM ROIs in the current study required the addition of FA. Previously, tractography FDi also showed strong correlations with FA (Roberts et al., 2005; Ukmar et al., 2012), supporting the relationship between these highly ranked features. The unique role of tissue alignment and orientation measures may facilitate early detection and even prediction of lesion pathology in future studies, promoting early management.

The core-shell analysis was designed to probe the intra-lesion patterns of pathology as seen in chronic MS lesions. When active, these lesions show demyelinated, hypocellular cores and inflammatory demyelinating shells; while inactive, present with hypocellularity with no active demyelination in lesion territories (Pittock and Lucchinetti, 2007). As such, the expected differences between core and shell are the degree of cellularity, where the hypocellular core should have higher diffusivity than the cell-rich, inflammatory shell. Indeed, the 3 DTI measures (MD, RD, and FA) ranked the highest in our core-shell analysis, and MD played a dominant role. While there is lack of core-shell studies using dMRI in the literature, prior evidence attests the sensitivity of MD to subtle changes in MS pathology. One report showed that MD increased 5 months before the occurrence of active MS lesions, and pre-lesion MD correlated significantly with the MD of the lesions 2 months after active inflammation (Liu et al., 2012). In the present study, the lack of significance of the MD model in AUROC compared to chance may be due to several reasons, including the small number of such lesions and their heterogeneity in pathology. For instance, chronic inactive lesions may not show significant core-shell differences, and their inclusion may have caused artifacts. Alternatively, robust analysis of core-shell activity in chronic active lesions can improve our understanding of the ongoing pathology and so disease progression in MS patients (Giorgio et al., 2014).

Our inter-patient analysis results appear to agree with the findings described above. Between patients who have high versus low disease activity, FA ranked the highest in lesion tissue classifications; while in NAWM classifications, ODF energy was the best. In general, T2 MRI lesion load reflects disease activity, and in a long-term, it correlates with patient disability in MS as seen after 20 years of disease onset (Fisniku et al., 2008). Therefore, patients with higher lesion load are expected to have greater tissue damage, likely affecting both neurite density and orientation, and so greater changes in lesion FA between the patient groups. On the other hand, increased disease activity would also implicate increased pathology in the lesion-free areas, such as the NAWM, yet the changes wherein should be much less than in the visible plaques. Indeed, our NAWM analysis suggests that the pattern of differences between high and low disease patients is mainly orientation based, as reflected by ODF energy, deserving further confirmation.

Collective observations of this study may suggest that MD and FA are important diffusion metrics for assessing MS pathology. In particular, they may serve as top options in core versus shell analysis, disease severity comparisons between patients (e.g., lesion FA), and at a lesser extent, lesion versus non-lesion analysis. However, parameters from the relatively complex models appear to be more competitive in NAWM-associated evaluations than the DTI indices, particularly the orientation-driven metrics (ODF energy and ODI), suggesting the value of multi-model analysis. Currently, dMRI is one of the unique imaging approaches that can potentially evaluate both myelin and axonal properties in vivo. However, a variety of other techniques are underway in this regard, including those for myelin mapping using either advanced or conventional MRI. For example, myelin estimation based on simultaneous T1/T2 relaxometry and proton density mapping correlated strongly with histological myelin as seen in post-mortem MS brain samples (Ouellette et al., 2020). Another study demonstrated that quantitative susceptibility provided additional myelin information in brain NAWM of MS patients, independent of FA and RD (Yu et al., 2019). Further, pre-operative myelin mapping using T1/T2 ratio showed the potential to predict outcomes of trigeminal neuralgia following Gamma knife radiosurgery, while FA and RD demonstrated similar values in this prediction as estimators of pre-operative axons (Li et al., 2020). These findings further support the importance of multi-dimensional analysis in MS and similar diseases, including combining multi-model dMRI, and various other candidate metrics of myelin and axons.

We note a few limitations in this study. The spatial resolution of our dMRI was moderate. However, we registered all diffusion outcomes to the high-resolution T1 MRI to mitigate the impact, and our ssHARDI maps appeared similar to those obtained using high resolution dMRI from the HCP both visually and quantitatively. Next, to optimize data quality, we excluded lesions that did not have a “clean” match of contralateral NAWM regions, and lesions not large enough for core-shell analysis, limiting the sample size. Nonetheless, SVM analyses enabled determination of critical parameters, so the risk of model overfitting is minimal. Finally, there were lesions that appeared overlapping and they were segmented as single confluent regions. Given the inhomogeneity of signal intensity in such lesions, it may have somewhat affected the lesion-NAWM results (e.g., underestimation of tissue differences). However, we applied ROI means in all quantitative analyses, which should have minimized the effect, if any. The impact on core-shell analysis is expected to be less than the above, as regions of inhomogeneity may be included in both lesion portions. In the future, we seek to use higher resolution DTI and evaluate patients with different types of diseases and lesion pathologies to confirm our findings, including datasets with histology-verified lesion activity. Additionally, we also plan to investigate the utility of the top-performing diffusion features in assessing pre-lesion NAWM pathologies, and in predicting the occurrence of MS lesions, to promote clinical applications.

Collectively, using commonly available dMRI data, it is possible to perform competitive ssHARDI modeling. Combining machine learning with robust ssHARDI and DTI metrics may provide advanced assessment of lesion and NAWM pathology, including mean diffusivity for intra-lesion pathology in MS and similar diseases.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by the Conjoint Health Research Ethics Board, the University of Calgary. The patients/participants provided their written informed consent to participate in this study.

OO participated in study design, data analysis and interpretation, and manuscript drafting and editing. W-QL participated in study design, data acquisition and analysis, and manuscript editing. BP, MK, and LM participated in study design, data acquisition, and manuscript editing. YZ participated in study design, data acquisition and interpretation, and manuscript drafting and editing. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.