Associations of increased interstitial fluid with vascular and neurodegenerative abnormalities in a memory clinic sample

The vascular and neurodegenerative processes related to clinical dementia cause cell loss which induces, amongst others, an increase in interstitial fluid (ISF). We assessed microvascular, parenchymal integrity, and a proxy of ISF volume alterations with intravoxel incoherent motion imaging in 21 healthy controls and 53 memory clinic patients - mainly affected by neurodegeneration (mild cognitive impairment, Alzheimer's disease dementia), vascular pathology (vascular cognitive impairment), and presumed to be without significant pathology (subjective cognitive decline). The microstructural components were quantified with spectral analysis using a non-negative least squares method. Linear regression was employed to investigate associations of these components with hippocampal and white matter hyperintensity (WMH) volumes. In the normal appearing white matter, a large fint (a proxy of ISF volume) was associated with a large WMH volume and low hippocampal volume. Likewise, a large fint value was associated with a lower hippocampal volume in the hippocampi. Large ISF volume (fint) was shown to be a prominent factor associated with both WMHs and neurodegenerative abnormalities in memory clinic patients and is argued to play a potential role in impaired glymphatic functioning.


Introduction
The clinical diagnosis of mixed type dementia assumes a spectrum, ranging from cognitive impairment primarily influenced by vascular pathology on one side, to Alzheimer's disease (AD) dementia mainly defined by neurodegeneration on the other side ( Emrani et al., 2020 ). Both vascular and AD-related neurodegen-Intravoxel incoherent motion (IVIM) provides the opportunity to estimate diffusion components on a microscopic level ( Le Bihan et al., 1986 ). Traditionally, the bi-exponential model is used to quantify two microstructural components from the IVIM signal: the microvascular pseudodiffusion and parenchymal diffusion ( Le Bihan et al., 1988 ;Le Bihan et al., 1986 ). The microvascular pseudodiffusion ( D mv ) represents microvascular blood flow within capillaries that approximates a diffusion process, which can subsequently be measured with diffusion imaging. The parenchymal diffusion ( D par ) represents the motion of water molecules within tissue and generally has a lower speed and amplitude ( Le Bihan et al., 1988 ).
The spectral analysis using the non-negative least squares (NNLS) method lays no constraints on the number of microscopic components that can be estimated by the model and has recently been used to identify a novel, intermediate diffusion component between the two traditional microstructural components ( De Luca et al., 2018 ;Keil et al., 2017 ;Whittall and MacKay, 1989 ;Wong et al., 2019 ). The amplitude of this intermediate component has been previously suggested as a proxy for (increased) ISF volume ( Wong et al., 2019 ). Thereby, IVIM allows for simultaneous investigation of the microvasculature, parenchymal microstructure, and ISF volume.
ISF plays a prominent role in the glymphatic system, a wasteclearance system which has previously been brought in relation to both neurodegeneration and vascular alterations in memory clinic patients ( Jessen et al., 2015 ;Mestre et al., 2017 ;Rasmussen et al., 2018 ;Rivera-Rivera et al., 2017 ). The ability to non-invasively measure ISF volume using spectral analysis, enables further exploration of pathologies associated with glymphatic impairment.
The current study aimed to investigate associations between microscopic diffusion and perfusion components and macroscopic MRI markers of neurodegenerative (i.e., hippocampal atrophy) and vascular (i.e., white matter hyperintensity (WMH) volume) pathology in a memory clinic sample. The applicability of IVIM will be investigated in a memory clinic sample with various degrees of vascular and neurodegenerative pathology. This study simultaneously investigates the microstructure and microvasculature with IVIM in memory clinic patients and healthy controls, covering the whole spectrum of vascular and neurodegenerative pathology, including individuals assumed to be mainly affected by neurodegeneration (mild cognitive impairment, AD dementia), vascular pathology (vascular cognitive impairment) and presumed to be without significant pathology (subjective cognitive decline and healthy controls).
We hypothesize that vascular pathology in terms of white matter hyperintensity volume is associated with increased ISF volume, microvascular changes and decreased parenchymal integrity in the NAWM, a region strongly influenced by vascular processes ( Rosenberg, 2009 ). Furthermore, decreased hippocampal parenchymal integrity and increased ISF volume are foreseen to relate to neurodegeneration in terms of a reduced hippocampal volume ( Moodley and Chan, 2014 ). The intermediate component is thought to be a proxy for ISF increases due to cellular alterations, and therefore is expected to be associated with both macrostructural vascular (WMH volume) and neurodegenerative markers (hippocampal volume) in the NAWM and hippocampi, respectively.

Participants
To obtain representative variation in WMH volume and hippocampal volume across the spectrum from normal cognition to dementia, cognitively normal older individuals as well as memory clinic patients with various degrees of cognitive impairment were included in this study ( Freeze et al., 2019 ). Neurodegenerative and vascular influences can be viewed as pathological dimensions with arbitrary cut-offs, which in reality show an overlap. Pure vascular dementia (dementia caused solely by vascular pathology) or pure AD dementia (solely caused by AD pathology) are uncommon in clinical practice ( Emrani et al., 2020 ;van der Flier et al., 2018 ). Therefore, we chose to analyze the memory clinic cohort as a whole, and not separately based on clinical diagnosis.
Patients were recruited at Maastricht University Medical Center and Zuyderland Medical Center. Inclusion criteria for patients were a clinical diagnosis of either subjective cognitive decline (SCD) ( Jessen et al., 2014 ), vascular cognitive impairment (VCI), mild cognitive impairment (MCI) ( Petersen, 2004 ) or clinical AD dementia ( McKhann et al., 2011 ), and mini-mental state examination (MMSE) score ≥ 20. Individuals were diagnosed with SCD when self-experienced cognitive decline was reported in comparison with a previously normal status and unrelated to an acute event, but no objective cognitive impairment was detected on any of the neuropsychological tests ( Jessen et al., 2014 ). Individuals were diagnosed with either MCI or VCI -dependent on the expected cause -when concern about cognitive functioning was reported by the patient or an informant, cognitive impairment was detected on at least one cognitive domain, and dementia was absent ( Petersen, 2004 ). Individuals were diagnosed with AD dementia if they met the NIA-AA core clinical criteria for AD dementia ( McKhann et al., 2011 ).
Control participants were recruited through means of advertisement in the local newspaper and online advertisement. Cognitively healthy controls had not visited a memory clinic previously, did not have any cognitive complaints or impairment in memory performance and had a MMSE score ≥ 27.
All participants were aged ≥ 55 years and did not show any structural abnormalities on MRI, nor had experienced recent (i.e., < 3 months before inclusion) ischemic or haemorrhagic stroke. Further details on inclusion and exclusion criteria are described in Freeze et al. (2019) .

Standard protocol approvals, registrations, and patient consents
Written informed consent was obtained from all participants prior to inclusion. The study protocol was approved by the Medical Ethics Review commission azM / UM (METC) in Maastricht, followed the ethical guidelines of the Dutch Medical Research Involving Human Subjects Act (WMO) and was in line with the Helsinki Declaration of Human Rights. This study was registered at Clinical-Trials.gov (Identifier: NCT02018913; date of registration: December 13, 2013).

MRI acquisition
All participants underwent a MRI protocol on the same 3.0 Tesla scanner with a 32-channel head coil (Philips, Achieva TX; Philips Healthcare, Best, the Netherlands), as previously described in more detail ( Freeze et al., 2019 ).

Image processing
Trace images were determined by calculating the average of all three diffusion-sensitive directions and corrected for head displacements, eddy current, and echo planar imaging distortions (Ex-ploreDTI version 4.8.4) ( Leemans et al., 2009 ). Resulting images were smoothed with a Gaussian kernel with a full width at half maximum of 3 mm (FSL version 6.0.1) ( Jenkinson et al., 2012 ).
Anatomical T 1 images were automatically segmented (FreeSurfer version 5.1.0), with visual inspection and manual adjustment when required ( Fischl, 2012 ). WMHs were delineated on the fluid-attenuated inversion recovery images using an in-house developed segmentation tool ( Jacobs et al., 2014 ) and manually checked, as previously described ( Freeze et al., 2019 ). From the segmented images, estimated intracranial volume (ICV), a normal appearing white matter (NAWM) mask, a hippocampi mask (bilateral), and the hippocampal volume (bilateral) were determined. As regions of interest, the hippocampi were chosen considering their prominent influence by neurodegeneration ( Fox et al., 1996 ;Moodley and Chan, 2014 ), and the total NAWM due to its strong influence by vascular processes (resulting in lesions such as WMHs) ( Rosenberg, 2009 ). Both the hippocampi and NAWM mask were coregistered and spatially down sampled to native IVIM space via the T 1 anatomical images (FLIRT, FSL) ( Jenkinson et al., 2002 ).

Spectral diffusion analysis
Spectral analysis using NNLS was conducted to analyze the IVIM data in a voxel-based manner, as previously described ( De Luca et al., 2018 ;Keil et al., 2017 ;Whittall and MacKay, 1989 ;Wong et al., 2019 ). Spectral analysis considers the signal to be comprised of a sum of multiple exponentially decaying components that fall within their own range of diffusion components. No constraints are laid on the number of components to be estimated by the model. Based on previous studies ( Wong et al., 2019 ), the spectrum was divided into three ranges: the parenchymal diffusion, 0.1 * 10 −3 < D < 1.5 * 10 −3 mm 2 /s; the intermediate diffusion component (proxy for ISF volume), 1.5 * 10 −3 ≤ D ≤ 4.0 * 10 −3 mm 2 /s; and the microvascular pseudodiffusion, 4.0 * 10 −3 < D < 10 0 0 * 10 −3 mm 2 /s ( Fig. 1 ). No regularization was applied to the spectrum.
For each voxel, the diffusion values were calculated for the microvascular pseudodiffusion ( D mv ), the intermediate diffusion ( D int ) and the parenchymal diffusion ( D par ) components. D mv , D int and D par were determined by selecting the position of D corresponding to the peaks in the spectrum, and subsequently taking the maxima of the peaks per component. A dictionary of D values was used ranging from 0.1 * 10 −3 to 10 0 0 * 10 −3 mm 2 /s with 200 logarithmically spaced values ( Wong et al., 2019 ). The amplitudes per diffusion compartment were defined as the sum of the calculated diffusion values' extension per compartment within the voxel. The least-square based model was constraint to non-negative values, as negative amplitudes have no physiological meaning ( Wong et al., 2019 ). The relative contribution of the intermediate component to the signal ( f int ) was determined by dividing the amplitude of spectral peaks within the intermediate diffusion range by the sum of the amplitudes of the peaks of the spectrum. The microvascular pseudodiffusion fraction ( f mv ) was determined in a similar manner, which means that the sum of f mv , f int and the remaining parenchymal diffusion fraction would result in a total of 100% per voxel.
The spectral analysis with NNLS resulted in the following diffusion parameters: the microvascular pseudodiffusion ( D mv ), which represents the diffusion value of the blood circulating in the microvasculature; the parenchymal diffusion ( D par ), which represents the diffusion value of the parenchyma; and the intermediate diffusion ( D int ), which represents the diffusion value of the intermediate component, presumed to be a proxy for ISF. The contribution of the amplitudes falling within each component's range were quantified as a fraction of the sum of all amplitudes in the spectrum. The microvascular pseudodiffusion fraction ( f mv ) represents the volume fraction of the microvascular signal, and the intermediate component fraction ( f int ) represents the volume fraction of the ISF component. Only voxels which contained a parenchymal peak were included in the further analysis, and when only an intermediate peak was present, this peak was categorized as falling into the parenchymal range. This correction was applied as it is physiologically unlikely that no parenchyma, but only ISF is present within the ROIs as defined in this study.
f mv and f int were corrected for the inversion pulse and T 1 and T 2 relaxation effects for each quantified component individually ( Wong et al., 2019 ). T 1 and T 2 relaxation properties of the microvascular and parenchymal compartment are retrieved from literature (T 1 = 1624 ms, T 2 = 275 ms and T 1 = 1081 ms, T 2 = 95 ms respectively) ( Lu et al., 2004 ;Simon et al., 2004 ;Wansapura et al., 1999 ). The T 1 times of the intermediate component were estimated using an iterative procedure and are voxel-specific, using initialization values of 30 0 0 ms and 1500 ms for T 1 and T 2 values respectively ( Wong et al., 2019 ). To assess whether any clinical group differences in estimated T 1 times of the intermediate component were present, an analysis of variance with Tukey post-hoc analyses was conducted. No significant group differences in estimated T 1 times were identified (Supplementary Table 1). The median values of D mv , f mv , D int , f int and D par were extracted for each parameter within the hippocampi and NAWM.
For comparison purposes, the bi-exponential model was applied to extract the traditional microvascular and parenchymal IVIM parameters ( Le Bihan et al., 1988 ), on which the statistical analyses were applied in a similar manner. The traditional bi-exponential analysis is restricted to the quantification of only two diffusion components (the microvascular pseudodiffusion and parenchymal diffusion), as the combination of only two exponential decay curves are fitted ( Le Bihan et al., 1988 ;Le Bihan et al., 1986 ).

Statistical analysis
Multivariable linear regression was employed to investigate the associations of WMH volume and hippocampal volume with the D mv , f mv , D int , f int and D par of the NAWM and hippocampi, while correcting for age, sex and ICV (IBM SPSS statistics version 25). One influential outlier was identified when investigating the association between hippocampal volume and f int in the NAWM (Cook's distance = 2.433) (Supplementary figure 1). However, statistical analyses including and excluding this particular case yielded no substantially different outcomes. Total WMH volume and hippocampal volume were assessed for normality using the Shapiro-Wilk Test. Total WMH volume was calculated per participant and subsequently log-transformed to better approximate a normal distribution. A threshold level of p < 0.05 was used to determine sta- tistical significance. Due to the exploratory nature of the current study, we did not correct for multiple comparisons.

Data availability
Anonymized data that support the findings of this study are available from the corresponding author, upon reasonable request from any qualified investigator.

Participant characteristics
Seventy-four individuals participated in this study, including patients with AD dementia (n = 15), MCI (n = 16), VCI (n = 9), SCD (n = 13) and healthy controls (n = 21). SCD and control participants were pooled together when displaying demographic characteristics (n = 34). Although the memory clinic cohort was analysed as a whole, sample characteristics including participant-and imaging characteristics and calculated IVIM measures are summarized per clinical group in Table 1 . Figure 2 A shows the various degrees of vascular and neurodegenerative pathology present in this cohort, by displaying the variation in WMH volume and hippocampal volume. This figure shows that in general, the control group seems to demonstrate a low WMH volume and high hippocampal volume, while the VCI group predominantly presents with a larger WMH volume and the MCI and AD groups show a lower hippocampal volume. However, no clear clinical cut-offs can be identified as the measures appear to overlap between the different groups, supporting the decision to analyze the memory clinic cohort as a whole, and not separately based on clinical diagnosis. In addition, representative example maps of f int are displayed ( Fig. 2 B-C), one of a VCI patient with a high WMH load and low hippocampal volume ( Fig. 2 B) and the other of a control participant with low WMH load and high hippocampal volume ( Fig. 2 C). The patient with VCI appears to have higher and more extensive f int values in the NAWM ( Fig. 2 B 1 ) and hippocampus ( Fig. 2 B 2 ) than the control participant ( Fig. 2 C 1 and C 2 respectively).

Associations of IVIM measures with WMH and hippocampal volume 3.2.1. Spectral analysis
The associations between WMH and hippocampal volume and the IVIM measures ( D mv , f mv , D int , f int and D par ) per ROI are provided in Table 2 . In the total sample, hippocampal volume was positively associated with D int ( β = 0.393, p = 0.001) and negatively associated with f int in NAWM ( β = -0.381, p = 0.001). Interestingly, the WMH volume showed the inverse pattern of associations with the same IVIM measures in NAWM ( Table 2 ), where a larger WMH volume related to a higher f int ( β = 0.406, p < 0.001) and a lower D int ( β = -0.270, p = 0.005). Moreover, hippocampal volume and D par of the NAWM showed a negative association. Within the hippocampi, hippocampal volume was negatively associated with f int ( β = -0.251, p = 0.032). No other IVIM measures showed any associations with WMH volume or hippocampal volume in any region ( p > 0.05).
The results did not change after including average absolute displacement as an additional covariate. Similarly, the correction for absolute (aSSR) and relative (rSSR) sum of squared residuals -accuracy measures of NNLS fit -did not alter the results (results not shown). This indicates that the identified associations were not attributable to variation in NNLS fit or subject motion.

The bi-exponential model
When using the traditional bi-exponential model ( Le Bihan et al., 1988 ), similar associations were found concerning the parenchymal diffusion component ( Table 3 ), as were found with the intermediate component using spectral analysis. Using the biexponential model, we observed an association of D par in the NAWM with WMH volume ( β = 0.296, p = 0.003) and hippocampal volume ( β = -0.301, p = 0.014) and in the hippocampi with WMH volume ( β = 0.193, p = 0.046) and hippocampal volume ( β = -0.517, p < 0.001).

Discussion
The current study investigated the relation of microvasculature, interstitial fluid, and parenchymal integrity with vascular (i.e., WMH volume) and neurodegenerative (i.e., hippocampal atrophy) biomarkers in memory clinic patients and controls. By evaluating associations between microscopic tissue alterations and common markers of neurodegenerative as well as vascular pathology (WMH volume) in this mixed study sample, we showed that a large ISF volume is significantly associated with a higher presence of vascular and neurodegenerative biomarkers.

A large intermediate component volume fraction in NAWM is associated with WMH volume and with hippocampal volume
We found associations of both WMH volume and hippocampal volume with f int in NAWM. A higher amplitude peak within the intermediate component is reflected by a higher f int and is in turn representative of a larger ISF volume in IVIM ( Wong et al., 2019 ). These findings suggest a greater ISF volume to be related to both vascular (WMH volume) and neurodegenerative (hippocampal volume) macroscopic alterations. This study investigated diffusion volume fractions in memory clinic patients, covering a broad spectrum of vascular and neurodegenerative pathology, includ-ing patients assumed to be principally affected by neurodegeneration (patients with AD and MCI), by vascular pathology (patients with VCI), and presumably without significant pathology (SCD and healthy controls). Although no previous diffusion studies have looked at ISF variations in a memory clinic sample which includes patients who are assumed to be principally affected by AD pathology, Wong et al. (2019) identified a relationship of high f int values with larger WMH volumes in a neurologic outpatient clinic patient sample affected by vascular pathology. In line with their results, our study identified the same positive association between f int values and WMH volume in the NAWM of a memory clinic sample. These findings suggest that f int could be a robust proxy for damage on the cellular level reflecting ISF changes, which in turn is associated with WMH (vascular) pathology. Furthermore, we identified that higher f int values in NAWM was related to hippocampal atrophy, a marker of neurodegenerative alterations.
In addition to the association of higher f int values with larger WMH and smaller hippocampal volume in the NAWM, we observed an association of a lower intermediate diffusion value ( D int ) in this ROI with both larger WMH and smaller hippocampal volume. When more tissue damage is present, a higher f int will be observed due to the increase in ISF volume related to pathophysiological processes such as neuroinflammation, atrophy and edema ( Abbott, 2004 ;Weller, 1998 ). Similarly, a lower D int can be observed, as more waste products will be present in the ISF, due to cellular breakdown and the potential presence of toxic proteins (e.g., amyloid beta and tau). Although tissue degradation itself might result in a larger ISF component (a higher f int ), the waste products in the extracellular matrix may hinder the diffusion of water molecules, leaving less free space for the water molecules in the ISF to diffuse, subsequently leading to lower intermediate diffusion values. The higher concentration of waste products in the extracellular matrix in damaged regions, will therefore lead to higher f int and lower D int values, regardless of the source of the parenchymal damage (i.e., vascular or neurodegenerative). The strong intrinsic relation between f int and D int might explain the similarity in magnitude of associations found with these parameters, although the associations of f int and D int with the MRI markers were found to be in opposite directions (i.e., negative versus positive). Thereby, this study demonstrated that ISF changes in memory clinic patients are not only related to macroscopic vascular abnormalities (WMH volume) but represent microstructural alterations in the NAWM which are linked to both underlying vascular and neurodegenerative abnormalities.

Hippocampal atrophy is associated with a large intermediate component volume fraction in the hippocampus
Interestingly, we observed f int in the hippocampi to be associated to the neurodegenerative marker (hippocampal volume), but not to the vascular marker (WMH volume). Consistent with these findings, neurodegenerative influences have been strongly established within this region and the hippocampi are known to be one of the first regions to be affected by atrophy in AD ( Fox et al., 1996 ;Moodley and Chan, 2014 ). Additionally, our observations are in line with previous results of Maier-Hein et al. (2015) , who used the free water (FW) imaging model in patients with MCI and AD dementia to determine a gradual increase in FW (representing the extracellular diffusion contribution) along the progression of AD, which significantly related to atrophy. In the same line of reasoning, there are studies that suggest the hippocampi to be less at risk of vascular alterations (such as WMH) than the NAWM, as a lower presence of WMH is found in the temporal lobe as compared to the other lobes ( Gootjes et al., 2004 ). This difference can be caused by the presence of a more robust vascular network and hippocampal vascular reserve in some patients and the absence of a specific vascular vulnerability of the hippocampi ( Gattringer et al., 2012 ;Perosa et al., 2020 ). For instance, the study of Perosa et al. (2020) shows that when a mixed vascularization of the hippocampi exists, fewer cSVD markers such as WMH are present. An increased vascular reserve of the hippocampi may be a protective factor for hippocampal structural integrity and cognitive function. While f int in the NAWM seems to be connected to both vascular (WMH volume) and neurodegenerative (hippocampal volume) abnormalities, as expected, hippocampal f int is only associated to neurodegenerative alterations. These findings support f int to be a proxy for damage on the cellular level reflecting ISF changes in the hippocampi, which are in turn associated with hippocampal atrophy and furthermore relate a large ISF volume to local macro-structural alterations.

WMH volume and hippocampal volume are not associated with the microvascular pseudodiffusion
The current study identified an association of WMH volume and hippocampal volume with f int but found no associations with f mv . The lack of microvascular associations in the current study contrasts with previous memory clinic studies, which did reveal connections between microvascular alterations and AD-related pathology. Multiple arterial spin labeling (ASL) studies have identified a hypoperfusion pattern including both the WM and hippocampi, decreasing along the continuum of AD ( Binnewijzend et al., 2016 ;Hays et al., 2016 ;Uh et al., 2010 ;van der Thiel et al., 2019 ;Zhang et al., 2017 ). Patients with AD dementia ( Binnewijzend et al., 2016 ;Zhang et al., 2017 ), patients with MCI ( Hays et al., 2016 ), and even individuals with subtle cognitive decline ( van der Thiel et al., 2019 ) have shown decreased perfusion in previous studies, thereby covering the whole spectrum of memory clinic patients. Although perfusion measures were generally observed to be lower in patients with MCI and AD dementia in WM regions ( Uh et al., 2010 ), hypoperfusion has also been indicated in the hippocampi . However, ASL-derived perfusion and IVIM microvascular pseudodiffusion represent different microvascular phenomena ( Le Bihan, 2019 ;Zhang et al., 2018 ), so caution needs to be taken when comparing our IVIM results with ASL studies. Regardless, these ASL-studies do suggest the occurrence of alterations at a microvascular level in memory clinic patients. ASL-measurements of perfusion might be influenced by variation in the blood supply to the brain ( Alsaedi et al., 2018 ;Okell et al., 2013 ;van der Thiel et al., 2018 ), while IVIM's microvascular pseudodiffusion gives a proxy of perfusion independent of variability in blood supply to the brain. Nevertheless, a previous IVIM study also supported the statement of alterations at a microvascular level in memory clinic patients, by demonstrating a decrease in microvascular pseudodiffusion in patients with VCI ( Wong et al., 2017 ).
Several factors can explain the absence of findings that connect vascular (WMH volume) and neurodegenerative (hippocampal volume) markers with the microvascular pseudodiffusion fraction in the hippocampi and NAWM in the current study. For instance, the difference in analysis methods could be responsible for the discrepancy in results concerning the microvasculature from our study and from previous studies. The added value of spectral analysis to analyze microstructural alterations in the brain is discussed in the following section.

The added value of spectral analysis
Our study implemented spectral analysis using NNLS in a memory clinic sample. Results of previous studies that have used the bi-exponential model, have indicated a relationship between f mv and vascular and neurodegenerative processes in the hippocampi ( van Bussel et al., 2015 ;Zhang et al., 2019 ). Similarly, when exploring the bi-exponential model, our memory clinic sample did show an association of hippocampal atrophy with the microvascular component in the hippocampi. Interestingly, when using spectral analysis, the current study did not find an association of hippocampal atrophy with the microvascular component in the hippocampi. Noteworthy, the association of the hippocampal ISF fraction ( f int ) with hippocampal atrophy found with spectral analysis is shown to be similar as the association of hippocampal atrophy with the hippocampal microvascular pseudodiffusion when the biexponential model is used. These findings suggest that when using bi-exponential models, the signal from the intermediate component could be leaking into the nearby microvascular component, biasing IVIM analysis when spectral analysis is not used. This may result in false associations of the microvasculature with macrostructural markers of pathology, which are actually caused by variations in ISF volume.
A previous study by Wong et al. (2017) mentioned enlarged perivascular spaces (PVS) as a potential underlying driver of this biasing effect. When water is flowing in a parallel direction to the vessels, as is the case in PVS, the signal from the PVS can falsely be contributed to the microvascular compartment ( Wong et al., 2017 ). In their study, a larger f mv was found to be associated with enlarged PVS in patients with vascular pathology, further supporting this notion ( Wong et al., 2017 ). Correspondingly, we were unable to find vascular (WMH volume) and neurodegenerative (hippocampal volume) associations with f mv , but did observe vascular and neurodegenerative associations with f int (ISF volume). An increased amount of ISF is argued to be present in enlarged PVS, where ISF can diffuse more freely ( Wong et al., 2019 ). Therefore, the enlargement of the PVS could underlie the observed large f int value (ISF) which is associated with WMH volume and hippocampal volume alterations in the current study.
Similarly, when using a bi-exponential model, associations can be identified of the parenchymal component with WMH and hippocampal volume in both the NAWM and hippocampi, while only an association between the parenchymal component in the NAWM and the hippocampal volume can be observed when the spectral analysis method is used. An impaired parenchyma ( D par ) estimated using the bi-exponential model, goes hand in hand with larger f int values in the same regions, as observed in the results of the current study using spectral analysis. When the parenchyma disintegrates, the remaining space will likely become more occupied by ISF ( Wong et al., 2019 ). Interestingly, increased extracellular fluid has previously been suggested to be the explanation for observed diffusion alterations which were assumed to indicate loss of parenchymal integrity in memory clinic patients with vascular pathology ( Duering et al., 2018 ). Our findings support this statement, and suggest a larger ISF volume to underly a decreased parenchymal diffusion component ( D par ) as also observed in memory clinic patients using bi-exponential models. With the current study we were able to simultaneously investigate the parenchymal integrity and ISF volume and subsequently suggest ISF alterations to underlie diffusion alterations that are associated with both WMH volume and hippocampal volume abnormalities.

The interpretation of the interstitial fluid fraction
Another potential explanation for the observed high f int values associated with WMH volume and hippocampal volume alterations could be the occurrence of microstructural tissue damage, such as neuronal or axonal degeneration and the subsequent loss of the cellular matrix and white matter fibre organization ( Maier-Hein et al., 2015 ). Within our memory clinic sample, microstructural tissue alterations are expected to be the leading factor associated with increased ISF volume, due to prominently observed cell loss within these patient groups ( Brun and Englund, 1981 ;Coleman and Flood, 1987 ). Therefore, we suggest the loss of tissue microstructure to underlie the observed large f int value, which is associated with WMH and hippocampal volume alterations in the current study.
The observed variations in ISF volume ( f int ) can be of influence on many mechanisms within the brain, such as the glymphatic system. The glymphatic system is a waste-clearance system which utilizes connecting PVS to efficiently eliminate soluble proteins and metabolites through the brains' ISF ( Iliff et al., 2012 ;Jessen et al., 2015 ;Mestre et al., 2017 ;Rasmussen et al., 2018 ). In line with the results of our study, glymphatic dysfunction has previously been brought in relation to both neurodegeneration and vascular alterations in memory clinic patients ( Jessen et al., 2015 ;Mestre et al., 2017 ;Rasmussen et al., 2018 ;Rivera-Rivera et al., 2017 ). An increase in pulsatility due to reduced vascular compliance has been found in patients with AD dementia, influencing the driving force of the system ( Jessen et al., 2015 ;Rivera-Rivera et al., 2017 ). Moreover, glymphatic failure renders the brain vulnerable to neurodegeneration and cognitive dysfunction ( Jessen et al., 2015 ). An important consequence of glymphatic dysfunction is that the waste clearance, including amyloid beta removal (A β; one of the main pathological hallmarks of AD), is attenuated ( Iliff et al., 2012 ). The potential to measure the ISF volume using spectral analysis with IVIM, opens possibilities to assess glymphatic impairment in memory clinic patients in a non-invasive manner.

Study considerations
Within our memory clinic sample, the individual diagnostic groups have a limited sample size, particularly the VCI group (n = 9). Due to the clinical and pathological heterogeneity present within the memory clinic sample ( Freeze et al., 2019 ), we chose to examine the study population as a whole using a continuous analytical approach. The mixed underlying pathology present in our study sample is also one of the main strengths of the study. Our sample covered a broad spectrum of vascular and neurodegenerative pathology, including patients assumed to be principally affected by neurodegeneration (patients with AD dementia and MCI), by vascular pathology (patients with VCI), and presumably without significant pathology (SCD and healthy controls), thereby enabling us to investigate both vascular and degenerative associations with microstructural alterations in cognitive decline. However, a limitation of this study is that AD biomarkers amyloid beta and tau were only known for 10 out of 15 patients with AD dementia (66.7%) and the presence of AD pathology was therefore not confirmed with amyloid biomarker evidence for all of them.
Furthermore, the aim of this study was to investigate associations between microscopic diffusion components and neurodegenerative (i.e., hippocampal atrophy) and vascular (i.e., WMH volume) MRI markers in a memory clinic sample. Although -like hippocampal atrophy for neurodegeneration ( Fox et al., 1996 ;Moodley and Chan, 2014 ;Scheltens et al., 1992 ) -WMHs are an established MRI marker of vascular pathology ( Freeze et al., 2019 ;Jacobs et al., 2014 ;Rosenberg, 2009 ;Wardlaw et al., 2013 ;Wong et al., 2019 ), these cannot be labeled as pure vascular markers, due to potential vascular influences leading to neurodegeneration and the interaction of these pathologic processes ( Jacobs et al., 2014 ). In addition, although histopathological studies have long ago indicated microvascular alterations -such as angiofibrosis and arteriosclerosis -to be associated with WMHs, WMHs can also be influenced by other -non-vascular -demyelinating processes such as Wallerian degeneration or edema ( Fazekas et al., 1991 ;Fazekas et al., 1993 ). Future studies are necessary to further investigate the complex interplay between neurodegenerative and vascular pathology and validate the specific place of ISF alterations within these pathologic processes.
The spectral analysis approach based on NNLS is a relatively new method and its limitations should also be considered. The ranges of the three components are assumed to be rigid over various brain regions. However, due to the anisotropy of white matter, the diffusivity of parenchymal component strongly varies over regions. For white matter regions with a relatively high fractional anisotropy, the parenchymal diffusivity could be higher than the upper boundary defined for the white matter, thereby leaking into the range defined for the intermediate component. In this study, we have tried to correct for this phenomenon by only including voxels that contain a parenchymal peak. Nonetheless, more research and further algorithm development is needed to account for the regional variability of component ranges.
Within our NNLS approach, a one-to-one relation between the exponential components as estimated with the NNLS and the microscopic compartments is assumed ( Wong et al., 2019 ). Another limitation of our method is that the T 1 and T 2 relaxation properties of the microvascular and parenchymal compartment are retrieved from literature and regionally invariant. However, the T 1 times of the intermediate component are estimated using an iterative procedure and are voxel-specific ( Wong et al., 2019 ). Future studies should further investigate the possibility to obtain quantitative region-specific estimates of T 1 and T 2 to improve correction for relaxation times.
In a previous study of Rydhög et al. (2017) , a threecompartment model was used to estimate a parenchymal, microvascular and FW compartment. In their study, the diffusion values per component were predefined based on values derived from the literature. However, as is stated in their discussion, the effective diffusivity of these compartments has been shown to vary in time and space ( Federau et al., 2014 ;Rydhög et al., 2017 ;Wirestam et al., 2001 ), and a variety of values are observed in the literature ( Federau et al., 2012 ;Federau et al., 2014 ;Nicolas et al., 2015 ;Wirestam et al., 2001 ). A potential explanation for the variation in diffusion values between studies could also lie in discrepancy in clinical groups, where slower of faster diffusivity might relate to specific disease states. The main advantage of spectral analysis using NNLS in the analysis of IVIM data, is that this approach is data driven. The diffusion values per compartment are determined by the model -within the prespecified diffusion ranges -on which the volume fractions are subsequently calculated. In our study, the diffusion values of ISF/FW ( D int ) were determined to lie between 2.3-2.6 mm 2 /s, However, in the discussed threecompartment model, these were predefined to be 3 mm 2 /s, leading to a potential overestimation of this diffusion value when not directly derived from the data. Within the spectral analysis approach, we leave space to estimate the exact diffusion values, which might be beneficial within clinical populations which represent with varying diffusivity within compartments.
In the current study, CSF suppression was applied to suppress the contamination of CSF, while maintaining the ISF signal due to the shorter T 1 time of ISF. Additionally, a potential bias of partial volume effects was minimized by taking the median value per ROI, to ensure that our statistical analysis would not be biased by outlier values. Thereby, we aimed to provide an accurate estimation of the ISF volume per ROI.
Prospective studies in larger patient groups are necessary to explore the potential future usage of the intermediate component as a biomarker for ISF volume associated with different clinical phenotypes of cognitive impairment or distinctive neurological disor-ders. These studies might be able to investigate the potential of f int as an early biomarker for microscopic damage associated with vascular and neurodegenerative pathology. Moreover, the clinical relevance of f int as a biomarker requires further studies using clinical outcomes.
Furthermore, it is important to note that granting the intermediate component has been previously validated as a potential measure for aberrant amounts of ISF in a patient sample ( Wong et al., 2019 ), the intermediate component as a proxy for ISF volume remains to be verified, for example in animal models, where tissue specimens can be validated microscopically. Although the usage of IVIM imaging seems to have great potential to gain more knowledge on the brains' microstructural composition in a memory clinic sample, prospective histopathological studies are needed to obtain more clarity on the physiological nature of the intermediate diffusion component and to further explore its relationship to glymphatic functioning. As previously suggested by Wong et al. (2019) , future studies should investigate the potential of IVIM imaging to detect glymphatic failure by combining IVIM using spectral analysis with measurements of cardiac pulsatility and aquaporin-4 dependent fluid movement.
Furthermore, the current study used three directions for IVIM data acquisition and analyzed the trace image to remove any effects of directionality on the IVIM parameters. However, incorporating direction dependent information of the IVIM diffusion signals might lead to interesting insights in relation to directionality of glymphatic flow patterns.

Conclusions
This study validated the applicability of IVIM in a memory clinic sample using spectral IVIM analysis and found associations between microstructural diffusion components and vascular (i.e., WMH volume) and neurodegenerative (i.e., hippocampal volume) MRI markers. We demonstrated that a larger intermediate component (suggested to be a proxy of ISF volume) in NAWM is associated with a larger WMH volume and a lower hippocampal volume. Moreover, we showed a relation between a high hippocampal intermediate component volume fraction and hippocampal atrophy. Thereby, we demonstrated prominent associations of a large intermediate component volume fraction with vascular and neurodegenerative abnormalities and obtained more pathophysiological insights by evaluating microscopic changes in a memory clinic sample influenced by both vascular and AD pathology and by linking them together. Furthermore, our observations might be related to impaired glymphatic functioning associated with the large ISF volume ( f int ) in memory clinic patients.

Verification
We state that the work described has not been published previously (except in the form of an abstract), that it is not under consideration for publication elsewhere, that its publication is approved by all authors and tacitly or explicitly by the responsible authorities where the work was carried out, and that, if accepted, it will not be published elsewhere in the same form, in English or in any other language, including electronically without the written consent of the copyright-holder.

Disclosure statement
All other authors declare that they have no conflict of interest.