Microstructural abnormalities of substantia nigra in Parkinson's disease: A neuromelanin sensitive MRI atlas based study

Abstract Microstructural changes associated with degeneration of dopaminergic neurons of the substantia nigra pars compacta (SNc) in Parkinson's disease (PD) have been studied using Diffusion Tensor Imaging (DTI). However, these studies show inconsistent results, mainly due to methodological variations in delineation of SNc. To mitigate this, our work aims to construct a probabilistic atlas of SNc based on a 3D Neuromelanin Sensitive MRI (NMS‐MRI) sequence and demonstrate its applicability to investigate microstructural changes on a large dataset of PD. Using manual segmentation and deformable registration we created a novel SNc atlas in the MNI space using NMS‐MRI sequences of 27 healthy controls (HC). We first quantitatively evaluated this atlas and then employed it to investigate the micro‐structural abnormalities in SNc using diffusion MRI from 133 patients with PD and 99 HCs. Our results demonstrated significant increase in diffusivity with no changes in anisotropy. In addition, we also observed an asymmetry of the diffusion metrics with a higher diffusivity and lower anisotropy in the left SNc than the right. Finally, a multivariate classifier based on SNc diffusion features could delineate patients with PD with an average accuracy of 71.7%. Overall, from this work we establish a normative baseline for the SNc region of interest using NMS‐MRI while the application on PD data emphasizes on the contribution of diffusivity measures rather than anisotropy of white matter in PD.


| INTRODUCTION
Parkinson's disease (PD) is a chronic, progressive disorder typically characterized by bradykinesia, rigidity, and tremor, and these symptoms have been implicated to degeneration of dopaminergic neurons in the substantia nigra pars compacta (SNc). These dopaminergic neurons contain neuromelanin, loss of which manifests as depigmentation of the SNc. This has been well established as an early histological feature of PD (Fearnley & Lees, 1991). Degeneration of these nigral, dopaminergic neurons may lead to alterations in microstructural organization of the regional gray matter, white matter and local myelination of the SNc in PD.
To gain understanding of the underlying microstructural changes in the SNc, studies have relied upon anisotropy and diffusivity measures computed from diffusion tensor magnetic resonance imaging (DT-MRI or DTI). Typically, these studies initially delineate the SNc and then perform analysis on the computed ROI. Table 1 provides a brief review of existing studies that provides the reported DTI findings in PD and is classified based on the technique used for SNc localization. Majority of these studies (Chan et al., 2007;Du et al., 2011;Knossalla et al., 2018;Langley et al., 2016;Loane et al., 2016;Peran et al., 2010;Rolheiser et al., 2011;Schwarz et al., 2013;Vaillancourt et al., 2009;Zhan et al., 2012) demonstrate a significant difference between PD and HC groups in at least one of the diffusion measures of SNc such as fractional anisotropy (FA), mean diffusivity (MD), radial diffusivity (RD) and axial diffusivity (AD). Contradictory to these findings, some studies (Aquino et al., 2014;Gattellaro et al., 2009;Menke et al., 2010; report no significant changes in any of these diffusion measures in PD. Taken together, the findings have been widely heterogeneous and perhaps could be implicated to inconsistencies in stages of disease severity as well as to variable sample sizes. More importantly, the techniques employed for delineating the SNc may have a significant bearing on the variability in the reported results. For example, studies have employed T2 weighted, proton density weighted spin echo and inversion recovery based contrasts to precisely localize the SNc ROIs (Atasoy et al., 2004;Duguid, De La Paz, & DeGroot, 1986;Geng, Li, & Zee, 2006;Hutchinson & Raff, 2000;Oikawa, Sasaki, Tamakawa, Ehara, & Tohyama, 2002;Pujol, Junque, Vendrell, Grau, & Capdevila, 1992;Stern, Braffman, Skolnick, Hurtig, & Grossman, 1989;Tuite, Mangia, & Michaeli, 2013). However, in a conventional T2 weighted MRI, there is variability in the hypo-intensity associated with the SNc as a result of increased iron deposition, in addition to the reduction in neuromelanin content (Deng, Wang, Yang, Li, & Yu, 2018;Langley et al., 2015;Langley et al., 2016;Wypijewska et al., 2010). This often leads to an inaccurate marking of SNc boundaries and a false pathological representation of dopamine degeneration (Langley et al., 2016). To obtain more precision in SNc delineation, recent studies have relied upon more sophisticated MRI protocols such as DTI where fiber tracking is employed to extract the SNc. (Menke et al., 2010;Sasaki et al., 2006;Zhang et al., 2015). However, this technique is highly dependent upon the choice of diffusion MRI protocol and the fiber tracking algorithm as well as is susceptible variations in manual fiber tracking. Nonetheless, direct visualization and segmentation of the SNc, is therefore, a simpler yet a precise option to ensure superior accuracy in analysis of PD. To this end, a novel MR NMS-MRI which is a 3 T T1-weighted high-resolution fast spinecho sequence is highly sensitive to the neuromelanin contained in the SNc and therefore renders the SNc as a hyper intense structure (Sasaki et al., 2006;Sasaki et al., 2008). This sequence is based on the paramagnetic properties of neuromelanin, a neuronal pigment which is a by-product of dopamine synthesis. Owing to the dopaminergic neuron loss in patients with PD, this normally hyper intense structure shows loss of normal signal intensity on NMS-MRI and therefore can be considered as a biomarker for PD. Multiple studies have demonstrated the clinical utility and accuracy of this sequence in patients with PD (Castellanos et al., 2015;Ohtsuka et al., 2013;Ohtsuka et al., 2014;Schwarz et al., 2011). To quantify these differences, the processing techniques that are currently used are based on visual inspection, or manual region of interest drawing followed by computation of volumes, contrast ratios or radiomics features and are arduous and time-consuming (Isaias et al., 2016;Kashihara, Shinya, & Higaki, 2011;Matsuura et al., 2013;Matsuura et al., 2016;Ogisu et al., 2013;Ohtsuka et al., 2014;Reimao et al., 2015;Sasaki et al., 2006;Schwarz et al., 2011;Shinde et al., 2019). To overcome this, NMS-MRI sequence can be employed to accurately localize and create a template of the SNc that can be utilized for analysis in parkinsonian disorders. Delineating the SNc in healthy controls and creating a template will not only offer better anatomical context to future studies but also provide a normative baseline and a ground-truth for comparison of multiple populations.
To this end, our work aims to generate a probabilistic atlas of the Another group of 27 healthy controls (Age = 38.67 ± 11.01, gender [M:F] =18/9) whose NMS-MRI sequence was acquired as part of a different study  of the same group, was used in the construction of our probabilistic atlas.

| Image acquisition
All subjects were scanned on a 3 T Philips Achieva MRI scanner using a 32-channel head coil. Diffusion weighted images (DWI) for these subjects were acquired using a single-shot spin-echo EPI sequence with rep-

| Atlas construction
Probabilistic atlas was built from NMS-MRI images of 27 subjects as shown in Figure 1. For each subject, bilateral substantia nigra ROIs (snROIs) were created from NMS-MRI scans by manual segmentation.
An author with expertise in the NMS-MRI sequence (rater1 [R1]-author SP), delineated the right and left SNc on the axial slices and created a 3D binary mask (snROIs) for all 27 subjects. The NMS-MRI images were then linearly registered to the T1 image of the same subject by performing affine transformation using FLIRT in FSL (Smith et al., 2004).
The T1 images of all subjects were preprocessed by performing motion correction, intensity inhomogeneity correction and skull stripping using Freesurfer 6.0 (Fischl, 2012) and were transformed to the MNI space, by employing a deformable registration using Advanced Normalization Tools (ANTS), wherein a symmetric diffeomorphic transformation model (SyN) was applied and optimized using mutual information. The SyN is a large deformation registration algorithm, which performs a bidirectional diffeomorphism and regularization using Gaussian smoothing of the velocity fields and has shown to outperform other nonlinear registration algorithms in preserving brain topology (Avants, Epstein, Grossman, & Gee, 2008). The transformations from NMS-MRI to T1 and from T1 to MNI were concatenated and were applied to the snROIs to transform them to MNI space. Along with visual inspection of each registered image, mean and variance of registered images was com-

| Quantitative validation of atlas
Two additional sets of snROI markings were generated, one was by manual segmentation (by rater 2 [R2]-author A.S) and the other was generated by a fully automated deep learning model known as U-Net (Ronneberger, P, & Brox, 2015) was used for segmentation of SNc. 2.5 | DTI preprocessing and analysis DWI images of patients with PD and healthy controls were manually visualized for quality assessment. All preprocessing steps were done using FSL5.0.9 (Smith et al., 2004) which included removing the nonbrain regions, correction for head movement and eddy current induced distortions using "eddy correct" tool that performs an affine transformation between baseline b0 image and gradient images. The resulting rotating parameters of the affine transformation were used to rotate the gradients back, to align them with the transformed images. Least square approximation method was implemented to reconstruct the diffusion tensor images using "dtifit", and the tensor fitting was checked for anatomical alignment. SNc were extracted for all subjects using MNI registered diffusion maps and the atlas described in earlier, which was thresholded at a 0.5 probability.  (Breiman, 2001). At each node of a tree, different subset of randomly selected predictors are considered, of which the best predictor is selected for further splits. Each tree is built using a different random bootstrap sample, which consists of approximately two-thirds of the total observations.

| Statistical analysis
In this study, RF model was implemented on a total dataset of 232 subjects, out of which 75% was used for training and 25% for testing. A total of eight normalized diffusion measures comprising AD, (c) At each step, features were ranked based on the order in which they were removed along with their relative feature importance.
This process was repeated 10 times to ensure stability in classification performance. The average accuracy, sensitivity and specificity and feature ranking from all repetitions were used to evaluate the model performance.

| Atlas construction and quantitative validation
The probabilistic and thresholded (0.5) atlas of the SNc computed from NMS-MRI images of 27 controls is shown in Figure 1.

| Asymmetry in diffusion measures of SNc
Asymmetry of all diffusion measures of bilateral SNc was observed for both HC and PD group, with PD group showing a higher significance as shown in Table 3

| Classification of PD and HC based on diffusion measures of SNc
The average classification accuracy, sensitivity and specificity for basic RF model were 73.4%, 0.736 ± 0.01, and 0.731 ± 0.01, respectively, whereas that for RF-RFE model was 71.7%, 0.736 ± 0.01, and 0.686 ± 0.05, respectively. Average feature ranking was consistent for basic RF and RF-RFE model wherein MD L , RD R and RD L were found to be the three topmost ranked features as shown in Table S1. ROC plots indicating average sensitivity and specificity performance of both classifier models is shown in Figure 4.

| DISCUSSION
We created a probabilistic atlas of the SNc by precisely extracting the SNc ROIs using NM rich MR sequence and employed it to accurately delineate the SNc to create a normative atlas that can be used in future PD studies. We applied this atlas to a large cohort of PD patients to gain understanding of the microstructural abnormalities.
Our results not only endorsed earlier findings but also facilitated fresh evidence supporting presence of micro-structural changes in PD Results of independent t test indicating differences in diffusion measures of bilateral substantia nigra in Parkinson's patient and healthy controls. Blue boxplot indicates healthy controls group whereas orange indicates PD patients group. NS: not significant, * p < .05, ** p < .005, *** p < .0005 substantia nigra compacta using diffusion MRI analysis. We demonstrated higher diffusivity values in the SNc in PD, with no changes in anisotropy and significant asymmetry of the diffusivity values.
The SN is anatomically divided into the SNc and the SN pars reticulata (SNr), where SNc is further subdivided into nigrosomes and the nigral matrix. Nigrosomes are the primary sub regions of the SNc where dopaminergic neurons are degenerated in PD (Blazejewska et al., 2013;Takahashi et al., 2018). The largest of these, nigrosome 1, is positioned in the lateral SN, and is most affected in PD. As described by Takahashi et al, the nigrosome is clearly a part of SNc and hence it is visualized on NMS-MRI sequence. TheT2 weighted images capture the elevated levels of iron mainly in SNr (Du et al., 2011;Langley et al., 2015;Langley et al., 2016;Langley, Huddleston, Sedlacik, Boelmans, & Hu, 2017). A study by Langley et al. demon-strated that T2-weighted and NMS-MRI are sensitive to different sub regions of SN (Langley et al., 2016), and the hypo-intensity observed on T2 images is unreliable in localizing SNc (Deng et al., 2018;Langley et al., 2016;Wypijewska et al., 2010). A recent study also demonstrated that increase in T2* weighted hypo intense signal is an indication of increase in iron deposition related to PD pathology (Langley et al., 2017). On similar lines, work by Visser et al. employed the FLASH sequence on 7 T MRI to delineate the substantia nigra. However, this sequence does not capture the SNc, as it is sensitive only to the elevated concentrations of ferritin that are prominently observed in SNr (Visser et al., 2016). Therefore, applying the SNr atlas to PD is inappropriate in understanding the abnormalities, which occur predominantly in the SNc owing to the dopaminergic neuronal loss in PD.
To alleviate these limitations, NMS-MRI has been employed to visualize and quantify the intensity contrast in SNc (Isaias et al., 2016;Matsuura et al., 2013;Matsuura et al., 2016;Ohtsuka et al., 2013;Ohtsuka et al., 2014;Reimao et al., 2015;Sasaki et al., 2006;Schwarz et al., 2011). Earlier studies on NMS MRI (Kitao et al., 2013;Sasaki et al., 2006), through post mortem analysis, have already demonstrated correlation between localization of SNc region from NMS MRI contrast and the histologically delineated SNc. Additionally, comparative study on SNc contrast sequences has shown that higher concentration of neuromelanin in SNc is captured by NMS-MRI (Langley et al., 2015). Moreover, these studies have corroborated its utility not only to render the SNc region but also as a volume or contrast ratiobased biomarker in PD (Isaias et al., 2016;Kashihara et al., 2011;Matsuura et al., 2013;Matsuura et al., 2016;Ogisu et al., 2013;Ohtsuka et al., 2014;Reimao et al., 2015;Sasaki et al., 2006;Schwarz et al., 2011). However, the techniques employed to extract and analyze the SNc are manual and time-consuming with low reproducibility.
Our work alleviated these issues by creating a SNc template which in future studies would be crucial to overcome the discrepancies in SNc localization by providing a normative baseline for comparison of results across studies.
Our atlas creation was based on uniform and accurate image registration of all subjects to the MNI space. A review study on 14 different T A B L E 3 FDR corrected (p-value = .05) results of independent t test for diffusion measures between healthy controls and patients with Parkinson's disease  nonlinear registration algorithms found that ART and SyN algorithms have consistently performed well across multiple datasets (Klein et al., 2009). We employed a symmetric diffeomorphic (SyN) registration using the ANTs toolbox for registering subject T1 images and subsequently the SNc masks, created from NM rich sequences of 27 subjects onto the MNI space as shown in Figure 1. The maximized optimization of space-time deformation maps in SyN and hierarchical interpolation performed in ANTs, increased normalization accuracy and preserved the brain topology, thus enhancing the registration precision of our probabilistic atlas (Avants et al., 2008;Klein et al., 2009). Each of the registrations was manually checked for precision in registration. The probabilistic atlas created, was thresholded at 50% probability, as it removed the voxels outside the expected SNc region as shown in Degree of myelination, axonal diameter and distance between extracellular membranes drive the changes in radial diffusivity, whereas diffusion anisotropy implies a directional alignment of white matter tracts (Beaulieu, 2002). Intuitively, the biological process of fiber disintegration and de-myelination which are associated with neurodegeneration, should lead to an increase in RD and reduction in FA values. However, neurodegeneration may involve multiple additional pathological processes such as changes in membrane permeability, restructuring of white matter fibers, glial alterations and damage to the intracellular compartment. The degree of variation in these processes may be contributing towards the proportional changes in diffusion tensors in all three dimensions, and thereby reducing the sensitivity of FA (Acosta-Cabronero, Williams, Pengas, & Nestor, 2010). Nevertheless, it is important to note that our diffusion protocol was limited to 15 gradient directions which may not facilitate the best model (Jones, Knosche, & Turner, 2013) for fiber tractography or connectivity, but is valid for computing diffusivity and anisotropy measures.
In concurrence with the clinical asymmetry typically reported in PD, which is implicated to an asymmetrical degeneration of dopaminergic nigral neurons, we observed significant asymmetry of diffusion measures in SNc. Although the FA was not significantly different between PD and controls, in patients with PD, we observed that the FA in left SNc was significantly lower when compared to the FA in right SNc (Figure 3). Similarly, MD, AD and RD also demonstrated significantly higher values in the left SNc when compared to right ( Figure 3). We did not have details pertaining to clinical laterality, owing to which we were unable to ascertain the concordance between clinical lateralization and diffusion asymmetry. However, earlier work by our group has illustrated the correlation of clinical asymmetry and laterality with the asymmetry of the SNc using contrast ratios on NMS-MRI .
Our correlation analysis did not demonstrate any significant associations of DTI measures with AoO, DoI, UPDRS-III OFF, or LEDD scores. However, a trend was observed between UPDRS, and FA, DoI and FA, MD, RD and between AoI and FA (Supplementary Figure S1).

| CONCLUSIONS
In conclusion, this study addressed a crucial question of uniform SNc

DATA AVAILABILITY STATEMENT
Anonymized data and code used in this manuscript will be shared at the request of qualified investigators. The SNc atlas constructed in this study is made freely available on github website.