In total, 178 subjects were included in this study. The data for 38 subjects were acquired at the Kiel University, Kiel (Germany). These subjects were divided into three groups: young healthy controls (HC_Y) (n = 13; mean age ± SD 30.63 ± 4.34 years), middle-aged healthy controls (HC_M) (n = 13; 53.15 ± 6.98 years) and 12 idiopathic PD patients (PD) [66.75 ± 7.18 years, mean Unified Parkinson’s Disease Rating Scale (UPDRSIII) score 34.5 ± 8.4_{[Med OFF]} and 14.30 ± 14.46_{[med ON]}, mean L-dopa equivalent dose 621.69 ± 394.85 mg, mean disease duration 13.6 ± 6.5, H and Y 3.8 ± 0.8_{[Med OFF]} and 2.7 ± 0.5_{[Med ON]}]. Demographic and group distribution details are shown in Table 1. For every participant, MRI was acquired using a 3 Tesla MRI scanner (Philips Achieva) with an 8 channel head coil. The whole-brain DTI scans were obtained with a voxel resolution of 2 × 2 × 2 mm³ covering a field of view of 224 × 224 × 120 mm. These diffusion-weighted images were acquired for 32 directions with a b value of 1000 s/mm³, echo time (TE) of 59 ms and repetition time (TR) of 11855 ms. Also, 60 contiguous slices with fat saturation “on” and 5 b₀ images were obtained for each acquisition. Additionally, high resolution T1 images were obtained using standard MPRAGE (magnetization-prepared 180 degrees radio-frequency pulses and rapid gradient-echo) sequence with TR = 7.7 ms, TE = 3.6 ms and flip angle = 8°. These scans consisted of 160 contiguous sagittal slices with a voxel resolution of 1 × 1 × 1 mm³ and a field of view of 224 × 224 × 170 mm. The study protocol used was approved by the local ethics committee in Kiel and all patients signed a written informed consent regarding the procedure.

As a control experiment, an additional 140 age- and sex-matched (with middle-aged healthy and PD group from Kiel dataset) healthy subjects were obtained from a publicly available dataset from the Nathan Kline Institute (NKI)¹ and fed into the same analytical framework. These healthy subjects were classified into two subgroups according to age as middle-aged healthy controls – NKI (HC_M–NKI) (n = 70; 53.71 ± 3.11 years) and elderly healthy controls – NKI (HC_E–NKI) (n = 70; 66.20 ± 2.53 years). Demographics are detailed in Table 1. In this cohort, whole-brain DTI scans were obtained using a 3T MRI scanner (Siemens Magnetom Trio Tim) with 32 channel head coil and 2 × 2 × 2 mm³ voxel resolution. For each subject, 64 contiguous diffusion weighted slices were obtained for 137 directions with b value of 1500 s/mm³ and TE and TR of 85 ms and 2400 ms, respectively. High resolution T1 images (176 slices) using standard MPRAGE sequence were also obtained for each subject with TE = 2.52 ms, TR = 1900 ms and voxel resolution of 1 × 1 × 1 mm³.

The MRI scans primarily obtained in DICOM format were converted into NIfTI format using the built-in MRIcron toolbox (Rorden, 2012). Then the obtained NIfTI images were pre-processed using the FDT tool in FSL (Behrens et al., 2003a,b, 2007). The complete processing protocol is explained in detail elsewhere (Groppa et al., 2014) and is depicted in Figure 1. In brief, by using FSL², microstructural analysis was performed in order to obtain FA, MD, AD, and RD. During pre-processing, images were first corrected for motion artifacts and eddy currents by applying an affine transformation, and then the brain extraction tool was used to remove the non-brain tissue (Smith, 2002).

For network analysis, the connectivity matrices were obtained for each subject using the seed masks for 116 ROIs defined by the Automated Anatomical Labeling atlas (Tzourio-Mazoyer et al., 2002). The links or the entries in the connectivity matrix represent the ratio between the number of samples (or streamlines) that pass through the ROI (j) and all generated streamlines from the ROI (i). This weighted connectivity index between the ROIs in the matrix was then analyzed using global and local network parameters obtained via the Brain Connectivity Toolbox (Rubinov and Sporns, 2010)³. The acquired undirected and weighted measures were used to compute the network parameters of distance, clustering and centrality to observe the influence over global information transfer, principle hubs, and reorganization for segregation or integration. The detailed formulation and definitions for the network parameters are explained elsewhere (Rubinov and Sporns, 2010). Finally, a two-sample Kolmogorov–Smirnov (KS) test followed by the Bonferroni–Holm (B–H) method for multiple comparison correction was performed to reveal significant (p < 0.05) differences in network measures between the groups.

Network-based statistic (NBS) analysis⁴ was used to assess differences in the interregional connectivity matrices between the groups. NBS analysis deals with the multiple comparisons problem posed by connectomic data by evaluating the null hypothesis at the level of interconnected subnetworks rather than individual connections (Zalesky et al., 2010). Here, the weighted undirected connectivity matrices obtained from probabilistic tractography were subjected to NBS analysis to identify the difference between the groups. A threshold of the t-statistic greater than 3 was determined using the effect size (Cohen’s d) and was used for individual connections and 5000 permutations were generated to build up the null distribution. Finally, the networks showing impaired connectivity between regions when compared to other groups obtained with significance of p < 0.05 were reported. Further details regarding the procedure are mentioned elsewhere (Korgaonkar et al., 2014).

In order to study the microstructural white matter changes, tract-based spatial statistics (TBSS) method was applied (Smith et al., 2006). The FMRIB58_FA_1 mm standard space image provided by FSL was used as the target for the nonlinear registration and the skeleton threshold was set to 0.2 (Woolrich et al., 2009; Jenkinson et al., 2012). For statistical analyses of the diffusion measures (FA, MD, RD, and AD), permutation test (using the randomize tool) (Winkler et al., 2014) was applied between the three subject groups to test six different contrasts (HC_Y >/< PD, HC_M >/< PD, HC_Y >/< HC_M). For the contrast HC_y-PD and HC_M-PD, age and sex were used as the nuisance variable and only sex for the contrast HC_y–HC_M. For each contrast, 5000 permutations were carried out with different thresholding options [voxel-based, cluster-based and threshold-free cluster enhancement (TFCE)]. Family-wise error (FWE, p < 0.05) correction was applied to account for multiple comparisons. In this study, TFCE was preferred over voxel-based thresholding and cluster-based thresholding due to its significant benefits of overcoming the problem of defining the cluster-forming threshold while keeping intact the sensitivity benefits (Smith and Nichols, 2009).

The compared groups did not differ significantly (p > 0.05) in terms of sex but differed significantly (p < 0.001) in terms of age.

No significant differences (p > 0.05) in network parameters were found between the two healthy control groups from the main cohort (HC_Y and HC_M). In the control analysis performed using the NKI dataset, we as well observed no significant (p > 0.05) differences between the two healthy control groups – HC_M–NKI (λ_mean = 0.0798 ± 0.0017, E_mean = 2.5771 ± 0.1023) and HC_E–NKI (λ_mean = 0.0805 ± 0.0029, E_mean = 2.5897 ± 0.1138) as shown in Figure 2.

From NBS analysis no significant differences were obtained from the comparison between the healthy control groups from either dataset used in the study.

Both characteristic path length and eccentricity were found to be significantly decreased (p = 0.0013 and p = 0.0242, respectively, B–H corrected) in PD patients (λ_mean = 0.068 ± 0.004, E_mean = 2.205 ± 0.114) when compared to HC_Y (λ_mean = 0.073 ± 0.002, E_mean = 2.401 ± 0.065) and HC_M (λ_mean = 0.072 ± 0.005, E_mean = 2.403 ± 0.16) as shown in Figure 2. Hence, confirming that the obtained network changes are specific to PD pathology and are independent of age.

Network-based statistic analysis (t-stat > 3, p < 0.05, corrected for multiple comparisons) revealed a network of anatomical regions with a significantly decreased connectivity profile in PD patients when compared to both healthy control groups. The analysis revealed 13 regions with significantly diminished connectivity strength in PD patients when compared with young healthy controls including bilateral superior and inferior frontal, anterior cingulate, olfactory cortex, gyrus rectus, insula, and left precentral and thalamus. Similarly, the comparison between middle-aged healthy controls (HC_M) and PD patients (HC_M > PD) generated 5 regions – bilateral inferior frontal, right cingulate gyrus, thalamus, insula, and left superior parietal region, as depicted in Figure 3.

The compared healthy control groups did not differ significantly (p > 0.05) in terms of sex. The comparison between the healthy subjects (HC_Y > HC_M) revealed only one significant cluster of FA difference in the right anterior corona radiata (Figure 4 and Table 2). No differences in AD, MD, or RD between the healthy control groups were identified. These findings highlight the differences in microstructural integrity occurring in healthy controls (as a consequence of physiological aging processes) in comparison to PD patients (as a consequence of the disease).

Group-wise FA comparison of PD patients and middle-aged healthy controls (PD < HC_M) showed significant (p < 0.05, FWE corrected, controlled for sex and age) differences in splenium, body and tapetum of corpus callosum, and fornix. For the comparison of PD and young healthy subjects (PD < HC_Y), the clusters were identified in splenium of corpus callosum, left corticospinal tract and corona radiata. All significant clusters are shown in Figure 4 and the detailed characteristics of the clusters are outlined in Table 2.

The analysis of AD showed significant clusters between PD patients and young healthy controls in the posterior limb of internal capsula (PD < HC_Y) and right external capsula (PD > HC_Y). Similarly, a significant group difference in MD was detected between PD patients and young healthy subjects (PD > HC_Y) in splenium and body of corpus callosum, external capsula, anterior, and superior corona radiata and fornix. Differences in MD values between the PD patients and middle-aged healthy controls (PD > HC_M) were identified in external capsula, anterior and superior corona radiata and splenium of corpus callosum. Furthermore, the group-wise RD analysis between PD and middle-aged healthy controls (PD > HC_M) identified significant differences predominantly in the body of corpus callosum and fornix and between PD patients and young healthy controls (PD > HC_Y) in external capsula, the body of corpus callosum, fornix and the posterior thalamic radiation. The clusters obtained in AD, MD and RD analysis are depicted in Figure 5.

The distance parameters evaluated here, namely characteristic path length and eccentricity, show the alterations in the structural network, which in turn provides valuable details about the information flows and network integration. In the case of characteristic path length, a measure of global integration, the shorter the path length the “more effective” the information flow (Sporns, 2011). Hence, the observed overall decrease in the presence of shorter paths within the networks in PD patients in comparison to healthy control groups clearly highlights the disruption of information flow within the networks. Moreover, this disruption was further supported by an overall decrease in eccentricity, an indicator of the importance of a region in the network, in the PD networks, reflecting an impediment to efficient network integration. We previously showed that the eccentricity in PD patients correlated inversely with the applied voltage during deep brain stimulation for optimal clinical response (Koirala et al., 2017). Hence, the eccentricity does not only differentiate the diseased group from healthy controls but also reflects the amount of network abnormality, accentuating its impact on network efficiency and stressing the overall importance of distance measures. Furthermore, in agreement with previous studies using functional MRI that show a decrease in global efficiency in PD patients in comparison to healthy controls (Skidmore et al., 2011) and also a notable loss of long-range connections in PD patients with mild cognitive impairment (Baggio et al., 2014), the observed decrease in the eccentricity and characteristic path length in PD patients in our study clearly indicates a disruption of global network integration and information transfer.

Network-based statistic analysis demonstrated specific networks of reduced connectivity mirroring the white matter structural abnormalities seen in PD patients when compared to healthy controls. In particular, the identified networks included the regions from motor and sensorimotor, basal ganglia and limbic systems that are primary involved in PD. As expected, the number of significantly different regions in the network comparison between young healthy subjects and PD patients was higher than for middle-aged healthy subjects and PD patients, corresponding to the microstructural alterations obtained from the same contrasts. Previous studies using NBS for PD patients showed significant differences in the connections primarily involved in key components of the limbic system, basal ganglia and sensorimotor areas (Nigro et al., 2016; Li et al., 2017). Here in this study, we further complement those findings by showing specific basal ganglia – frontal networks when compared to healthy controls.

Previous functional MRI studies on the reorganization of brain networks in aging have shown that the default mode network is the most compromised one during the process of aging (Jie et al., 2014; Sala-Llonch et al., 2015). Hence, the anatomical regions found in the network when comparing young healthy subjects and PD patients shows some overlapping regions for aging, which vanish when compared with the middle-aged healthy controls. In addition, the overlap of the brain networks during development and disorders has been previously discussed (Douaud et al., 2014), which further highlights the importance of network metrics for the differentiation between aging- or PD-specific processes. The structural abnormalities occurring in PD patients in frontal, parietal and cingulate cortex have been shown using both white matter (Vercruysse et al., 2015) and gray matter analysis (Beyer et al., 2007; Rektorova et al., 2014; Mak et al., 2015). Hence, the differences in connectivity properties obtained from these regions using NBS analysis indicate that the white matter networks play an important pathophysiological role in PD. Moreover, no differences were identified in the comparison between the healthy subject groups in either cohort, which strongly demonstrates that the detected network changes are not the result of aging but rather derive from PD pathology.

Several previous studies reported that the reduced FA in olfactory cortex and gyrus rectus explains the dysfunction within the olfactory system and white matter tracts connecting it to other brain regions, causing olfactory impairment that is present even in the premotor phase of the disease (Ibarretxe-Bilbao et al., 2010; Doty, 2012). The insula is known to be an integrating node for multiple brain networks and its involvement in PD, particularly in relation to the non-motor symptoms, was repeatedly stipulated (Christopher et al., 2014; Criaud et al., 2016). Although diffusivity properties obtained from DTI are indirect measures of white matter integrity, it could be speculated that the PD-related neurodegenerative mechanisms (e.g., alpha-synuclein-driven damage to the presynaptic terminals) may impair axonal transport and consequently produce axonal degeneration (Nigro et al., 2016). The microstructural analysis performed in our study revealed white matter alterations linked to the disrupted network topology in PD patients. Lower FA values were identified in PD patients in the body, splenium and tapetum of the corpus callosum, fornix, and corona radiata – the fiber tracts known from previous studies to be affected in PD and considered primary substrates of cognitive impairment (Matsui et al., 2006; Zheng et al., 2014). These fiber tracts were also detected in the networks obtained from the NBS analysis, further indicating that the observed network alterations potentially cause the disconnection phenomena within the white matter.

Moreover, the significant alterations in MD observed between PD patients and healthy controls are in line with previous studies addressing the diffusion alterations associated with PD (Hall et al., 2016). An increase in RD is an indicator of increased demyelination and axonal degeneration (Alexander et al., 2011), and it has been shown that in PD the unmyelinated axons are more vulnerable to degeneration than myelinated axons (Orimo et al., 2011). Hence, the clusters of higher RD values in PD patients could bring axonal degeneration in PD patients to the foreground.

There were only minimal white matter changes between the young and middle-aged healthy control groups, which could be the result of insignificant variations in the white matter volume occurring between the ages of 30–55 (Sowell et al., 2003), the age interval of the healthy controls from our study. More importantly, not all of the regions of FA differences, detected from the comparison between PD patients and healthy controls were identified between healthy subject groups, which further reinforces that microstructural patterns of aging and neurodegeneration are distinct. It is known that AD alterations relate to the variation in axonal integrity within different brain regions during the life span (Kumar et al., 2013). Here, the differences in AD in PD patients in comparison to young healthy subjects were detected in more regions than shown in Kumar et al. (2013) for physiological aging. Thus, we further speculate that the decrease in AD might indicate neurodegeneration while an increase reflects compensation.

Overall, we found that network distance parameters serve as a novel approach for a non-invasive investigation of the emerging PD pathology. The performed analyses are substantiated by regional microstructural alterations showing an involvement at the anatomical level with predilection sites for the impact of PD pathology. The differential network reorganization pattern in PD compared to physiological aging facilitates the understanding of network modifications in PD pathophysiology. The primary motivation of this study was to provide the distinction between the network changes during aging and due to PD in order to demonstrate specific neurodegenerative network and microstructural alterations that occur in PD, which are beyond physiological aging. Hence, the use of different age groups in the study strengthens this objective. In one limitation of our study, we used the aged matched healthy controls dataset obtained from an open source, to control for the network changes due to physiological aging. As a result, the mean network measures obtained for two cohorts differ marginally but consistently (see Figure 2). As shown in the previous studies the connectivity values based on probabilistic tractography vary upon DTI directions (Vaessen et al., 2010), different scanner (Owen et al., 2013), applied density threshold (Andreotti et al., 2014) or different approach of obtaining the measures (Bonilha et al., 2015). We assume that the difference in our study is also due to the difference in DWI sequence and the scanning site. Furthermore, it would be very interesting to see the relationship with these network outcomes and structural changes with the neuropsychological data to associate these network topological disruptions with clinical outcomes. However, for the clinical implementation of the studied methodology, future studies with higher sample size, homogeneous data acquisition and correlation to the clinical parameters should be investigated.