Axonal degeneration as substrate of fractional anisotropy abnormalities in multiple sclerosis cortex

Abstract

Cortical microstructural abnormalities are associated with clinical and cognitive deterioration in multiple sclerosis. Using diffusion tensor MRI, a higher fractional anisotropy has been found in cortical lesions versus normal-appearing cortex in multiple sclerosis. The pathological substrates of this finding have yet to be definitively elucidated. By performing a combined post-mortem diffusion tensor MRI and histopathology study, we aimed to define the histopathological substrates of diffusivity abnormalities in multiple sclerosis cortex. Sixteen subjects with multiple sclerosis and 10 age- and sex-matched non-neurological control donors underwent post-mortem in situ at 3 T MRI, followed by brain dissection. One hundred and ten paraffin-embedded tissue blocks (54 from multiple sclerosis patients, 56 from non-neurological controls) were matched to the diffusion tensor sequence to obtain regional diffusivity measures. Using immunohistochemistry and silver staining, cortical density of myelin, microglia, astrocytes and axons, and density and volume of neurons and glial cells were evaluated. Correlates of diffusivity abnormalities with histological markers were assessed through linear mixed-effects models. Cortical lesions (77% subpial) were found in 27/54 (50%) multiple sclerosis cortical regions. Multiple sclerosis normal-appearing cortex had a significantly lower fractional anisotropy compared to cortex from non-neurological controls (P = 0.047), whereas fractional anisotropy in demyelinated cortex was significantly higher than in multiple sclerosis normal-appearing cortex (P = 0.012) but not different from non-neurological control cortex (P = 0.420). Compared to non-neurological control cortex, both multiple sclerosis normal-appearing and demyelinated cortices showed a lower density of axons perpendicular to the cortical surface (P = 0.012 for both) and of total axons (parallel and perpendicular to cortical surface) (P = 0.028 and 0.012). In multiple sclerosis, demyelinated cortex had a lower density of myelin (P = 0.004), parallel (P = 0.018) and total axons (P = 0.029) versus normal-appearing cortex. Regarding the pathological substrate, in non-neurological controls, cortical fractional anisotropy was positively associated with density of perpendicular, parallel, and total axons (P = 0.031 for all). In multiple sclerosis, normal-appearing cortex fractional anisotropy was positively associated with perpendicular and total axon density (P = 0.031 for both), while associations with myelin, glial and total cells and parallel axons did not survive multiple comparison correction. Demyelinated cortex fractional anisotropy was positively associated with density of neurons, and total cells and negatively with microglia density, without surviving multiple comparison correction. Our results suggest that a reduction of perpendicular axons in normal-appearing cortex and of both perpendicular and parallel axons in demyelinated cortex may underlie the substrate influencing cortical microstructural coherence and being responsible for the different patterns of fractional anisotropy changes occurring in multiple sclerosis cortex.

diffusion tensor MRI, multiple sclerosis, histopathology, cortical lesions, neurodegeneration

See Ropele and Fazekas (doi:10.1093/brain/awz160) for a scientific commentary on this article.

Introduction

Although multiple sclerosis has classically been defined as an inflammatory-demyelinating white matter disorder, several pathological and MRI studies have provided evidence for extensive cortical involvement, in terms of focal lesions, microstructural tissue abnormalities and irreversible tissue loss (Geurts and Barkhof, 2008; Calabrese et al., 2010a; Filippi et al., 2017, 2019; Magliozzi et al., 2018).

Pathological and technical issues make imaging of multiple sclerosis cortical lesions challenging. These include their small size, their frequent location in superficial cortical layers determining partial volume effects with the CSF and their poor contrast with the surrounding normal-appearing cortex.

However, several MRI studies have demonstrated that cortical lesions are present in all multiple sclerosis clinical phenotypes (Calabrese et al., 2010a), have a role in diagnostic work-up of suspected multiple sclerosis (Filippi et al., 2018; Preziosa et al., 2018), and are associated with locomotor and cognitive dysfunctions (Roosendaal et al., 2009; Calabrese et al., 2010b; Scalfari et al., 2018).

Recently, several advanced MRI techniques, including diffusion tensor and magnetization transfer MRI (Filippi et al., 2017), neurite orientation dispersion and density imaging (Granberg et al., 2017), and other quantitative MRI techniques (i.e. T₁- and T₂-relaxomentry, R2*, and T₁/T₂ ratio) (Gracien et al., 2016a, b; Jonkman et al., 2016a; Nakamura et al., 2017; Righart et al., 2017) have been applied to investigate grey matter damage due to their higher sensitivity and specificity to pathological changes. Among them, diffusion tensor MRI is contributing to better characterize the pathological substrates occurring in multiple sclerosis normal-appearing cortex and cortical lesions and to monitor disease evolution (Calabrese et al., 2011; Filippi et al., 2013, 2017).

At present, only a few in vivo studies have applied diffusion tensor MRI (Poonawalla et al., 2008; Calabrese et al., 2011; Filippi et al., 2013; Yaldizli et al., 2016; Granberg et al., 2017; Preziosa et al., 2017) to grade the severity of intrinsic cortical lesion damage and its clinical relevance in multiple sclerosis. Interestingly, while normal-appearing cortex is typically characterized by a decreased fractional anisotropy and increased mean diffusivity compared to healthy controls (Vrenken et al., 2006; Poonawalla et al., 2008; Filippi et al., 2013; Preziosa et al., 2017), an unexpected but consistent finding is that cortical lesions have an increased fractional anisotropy compared to normal-appearing cortex (Poonawalla et al., 2008; Calabrese et al., 2011; Filippi et al., 2013; Yaldizli et al., 2016; Preziosa et al., 2017). Interestingly, mean diffusivity values in cortical lesions were found to be either increased (Poonawalla et al., 2008; Calabrese et al., 2011), not significantly different (Filippi et al., 2013; Granberg et al., 2017; Preziosa et al., 2017) or decreased (Yaldizli et al., 2016) compared to normal-appearing cortex. It is noteworthy that cortical diffusion tensor MRI changes are clinically relevant, since they correlate with clinical disability (Calabrese et al., 2011; Yaldizli et al., 2016) and differentiate multiple sclerosis clinical phenotypes (Filippi et al., 2013).

Despite the steady progress of in vivo results, the pathological substrates underlying these diffusion tensor MRI-derived microstructural abnormalities have yet to be definitively elucidated. Different mechanisms (not necessarily mutually exclusive) have been proposed to contribute to such fractional anisotropy changes (Poonawalla et al., 2008; Calabrese et al., 2011; Filippi et al., 2013; Preziosa et al., 2017). These include the presence of demyelination, activated microglia, and neurodegeneration. Only one post-mortem MRI and histopathology correlation study directly compared fractional anisotropy changes and histopathological markers, and suggested an increased cellular density and tissue compaction as possible substrate (Jonkman et al., 2016b). However, a more detailed analysis is needed, since that study mainly investigated the role of activated microglia in determining fractional anisotropy changes in multiple sclerosis cortex (Jonkman et al., 2016b).

A better characterization of the pathophysiological mechanisms determining the cortical microstructural abnormalities measured using post-mortem MRI might allow one to identify new specific and reliable markers to monitor disease progression. Accordingly, we aimed to better define the pathological substrates responsible for diffusion tensor MRI-derived microstructural tissue abnormalities both in multiple sclerosis normal-appearing cortex and in cortex showing cortical lesions by combining a pipeline including post-mortem in situ diffusion tensor MRI followed by a comprehensive histopathological assessment.

Materials and methods

Donors included in the present study

All multiple sclerosis donors were registered and collected by the Netherlands Brain Bank, Amsterdam, the Netherlands (http://www.brainbank.nl/), while all non-neurological controls were collected by the Normal Aging Brain Collection Amsterdam (NABCA; http://nabca.eu) (Jonkman et al., 2019) and were registered with the body request program at the department of Anatomy and Neurosciences, VU University medical center, Amsterdam, the Netherlands. All donors gave written informed consent for the use of their tissue and medical records for research purposes. Permission for performing MRI, autopsies, use of tissue, and access to medical records was granted by the institutional ethics review board.

From November 2013 to December 2017, MRI and pathological data were collected from 16 patients with clinically and neuro-pathologically confirmed multiple sclerosis and 10 age- and sex-matched non-neurological controls with no pathological findings suggestive of neurological disorders who underwent whole-brain in situ (brain still in the skull) MRI acquisition followed by autopsy (Table 1) as previously described (Popescu et al., 2015,2016; Jonkman et al., 2019).

Table 1

Main demographic and MRI findings in multiple sclerosis patients and non-neurological controls

Variable	Non-neurological controls n = 10	Multiple sclerosis patients n = 16	P-value
Male / female	5 / 5	5 / 11	0.425a
Median age, years (IQR)	73.0 (70.5,77.3)	64.0 (53.6,76.8)	0.140b
Median disease duration, years (IQR)	–	31.5 (24.3,41.8)	–
Clinical phenotype (primary / secondary progressive)	–	4 / 12	–
Median post-mortem delay, min (IQR)	613 (420,724)	495 (460,535)	0.162b
Mean T2 white matter lesion volume, ml (SD)	8.7 (8.2)	47.5 (24.5)	<0.001c
Mean normalized brain volume, ml (SD)	1464 (53)	1386 (77)	0.013c
Mean grey matter volume, ml (SD)	738 (41)	712 (51)	0.032c
Mean cortical volume, ml (SD)	575 (32)	564 (39)	0.166c
Mean white matter volume, ml (SD)	726 (30)	677 (47)	0.034c
Mean cortical FA (SD)	0.229 (0.018)	0.219 (0.010)	0.203c
Mean cortical MD, mm2s−1 × 10−3 (SD)	0.464 (0.034)	0.545 (0.055)	0.010c
Mean cortical AD, mm2s−1 × 10−3 (SD)	0.561 (0.046)	0.653 (0.068)	0.008c
Mean cortical RD, mm2s−1 × 10−3 (SD)	0.437 (0.037)	0.510 (0.057)	0.009c

^aFisher exact test; ^bMann-Whitney U-test; ^clinear models adjusted for sex, age and post-mortem delay.

Note: Fractional anisotropy is a dimensionless quantity; see main text for further details.

AD = axial diffusivity; FA = fractional anisotropy; IQR = interquartile range; MD = mean diffusivity; RD = radial diffusivity; SD = standard deviation.

MRI acquisition

The MRI protocol comprised whole-brain in situ MRI sequences, as described before (Popescu et al., 2015, 2016; Jonkman et al., 2019). Briefly, after arrival in the hospital, each subject underwent an MRI scan with the brain in situ, similar to in vivo MRI scanning. Using a 3 T whole-body scanner (Signa-MR750, General Electric Medical Systems) and an eight-channel phased-array head-coil, the following sequences were acquired from the brain in situ: (i) sagittal 3D fluid attenuation inversion recovery (FLAIR); (ii) sagittal 3D T₁-weighted fast spoiled gradient echo sequence; and (iii) axial 2D echo-planar imaging (EPI) with diffusion gradients applied in 30 non-collinear directions and two optimized b factors (b₁ = 0 and b₂ = 900 s/mm²) (see Supplementary material for details).

Autopsy procedure and tissue selection

A detailed description of our data collection pipeline can be found elsewhere (Seewann et al., 2009; Jonkman et al., 2019). Briefly, after whole-brain in situ MRI acquisition, the brain was removed from the skull and cut into 10-mm thick coronal brain slices. From these slices, tissue blocks were obtained from six standardized anatomical regions: the superior and inferior frontal gyrus, the anterior and posterior cingulate cortex, the superior temporal gyrus and the inferior parietal lobule. See Supplementary material and Supplementary Table 1 for details.

Histopathology processing and analysis

Tissue blocks were fixed in 10% buffered formalin and embedded in paraffin (Seewann et al., 2009). Paraffin tissue blocks were cut to obtain 10- or 20-µm thick sections which were mounted on SuperFrost Plus™ glass slides (VWR international). Sections were deparaffinized in xylene and rehydrated in graded series of ethanol and rinsed in Milli-Q® water. Sections were then dried vertically overnight at room temperature and subsequently dried horizontally for 3 h at 45°C and then overnight at 37°C.

Consecutive 10-µm thick sections were stained for myelin [anti-proteolipid protein (PLP)], microglia/macrophages [anti-ionized calcium-binding adapter molecule 1 (Iba-1)], astrocytes [glial fibrillary acidic protein (GFAP)] and for axons (Bielschowsky silver staining). A further 20-µm thick section was stained for Nissl (thionin; Thermo Fisher Scientific), for visualization of all cells and anti-neuronal nuclear antigen (NeuN), to identify neurons. One section per tissue block was evaluated for each staining. Staining procedures details can be found in the Supplementary material.

From each tissue section, based on PLP staining, cortical lesions were identified as cortical areas with a complete demyelination; lack of PLP). Then, regions of interest of normal-appearing cortex (without demyelination) and of demyelinated cortex (cortical areas showing cortical lesions) were defined. All the defined regions of interest included the entire width of the cortex, since the acquired MRI sequences did not allow a proper cortical sub-parcellation. At least two regions of interest, covering together ∼16 mm² of cortex, were identified per tissue section at ×25 magnification using a DM/RBE photomicroscope (Leica).

Regions of interest with the same topography and similar shape, were also defined for microglia, astrocyte, neuronal and axonal assessments on consecutive sections of the same regions to be as similar as possible to those defined for myelin evaluation.

Quantification of cortical myelin density

Images of myelin stained sections (PLP) were digitized using a Leica DM/RBE photomicroscope at ×50 magnification attached to a minimal clinically important difference (MCID) elite image analysis system (Imaging Research Inc). Images were stitched to obtain a single image of the entire section.

ImageJ (Schindelin et al., 2012) (version 1.52a, https://imagej.net/Fiji) was used to superimpose the regions of interest onto the images. After computing a mask of the myelin staining for each region of interest, cortical myelin density was expressed as the percentage of total area of region of interest stained for myelin (Fig. 1A).

Figure 1

Pathological substrates evaluated in normal-appearing and demyelinated cortices. (A) Myelin density was quantified from images of proteolipid protein (PLP) stained sections using ImageJ (Schindelin et al., 2012) (version 1.52a, https://imagej.net/Fiji). After ImageJ processing, we obtained a binarized mask of myelin and its density was defined as the ratio of black pixels (representing myelin) divided by the total amount of pixels included in the defined region of interest (ROI). (B) Microglia density was quantified from images of anti-ionized calcium-binding adapter molecule 1 (Iba-1) stained sections using Nuance spectral imaging device and associated software (Nuance version 3.0.2). (C) Astrocyte density was quantified from images of GFAP-stained sections using ImageJ. After ImageJ processing, we obtained a binarized mask of myelin and its density was defined as the ratio of black pixels (representing astrocytes) divided by the total amount of pixels included in the defined region of interest. (D) Neuron density and volume and total cell density were quantified from images of anti-neuronal nuclear antigen (NeuN) and Nissl stained sections using stereology (Stereo Investigator software). (E) Axon density and orientation were quantified from images of Bielschowsky silver stained sections using ImageJ. See main text for further details.

In regions of interest with demyelinated cortex, cortical lesions were classified according to their location within the cortex (Bo et al., 2003) as type I (cortico-subcortical lesions affecting both grey matter and white matter, type II (small perivenous intracortical lesions not affecting white matter or the pial surface), type III (extending inward from the subpial cortical layers), and type IV (extending through the whole cortical width but without passing its border with the white matter).

Quantification of cortical microglia density

Images of Iba-1 stained sections (microglia) were acquired and analysed using a Leica Ctr 5000 microscope (Leica Microsystems) with a Nuance spectral imaging device and associated Nuance spectral imaging software (Nuance version 3.0.2, Caliper Kife Sciences, Inc, a Perkin Elmer Company, Hopkinton, MA). Different images were acquired including cortical layers I–II, layer III and layers IV–VI, respectively, at ×200 magnification. To obtain more reliable measures of microglia density, two images were acquired per layer, resulting in six images per region of interest and at least 12 per section. Based on the spectrum of Iba-1 immunoreactivity, a mask was obtained and microglia density was expressed as percentage of the total image stained for Iba-1 per mm² (Fig. 1B).

Quantification of cortical astrocyte density

Images of GFAP stained sections (astrocytes) were digitized using a Vectra Polaris Automated Quantitative Pathology Imaging System (Perkin Elmer) at ×200 magnification. ImageJ (Schindelin et al., 2012) (version 1.52a, https://imagej.net/Fiji) was used to superimpose the regions of interest, based on the myelin staining, onto the images. After computing a mask of the astrocyte staining for each region of interest, cortical astrocyte density was expressed as the percentage of total area of region of interest stained for astrocytes (Fig. 1C).

Quantification of cortical neuron density and volume and total cell density

To quantify total cell density and neuron density and volume, a Leica DMR microscope (Leica Microsystems) in combination with Stereo Investigator software (MBF biosciences, Williston, VT) was used for well-established procedures of cell counting and volume measurements. In particular, to obtain density measures unbiased for orientation, shape and volume of the neurons and glia cells, the optical fractionator tool (Gundersen and Jensen, 1987) was used and applied on 20-µm thick sections (Fig. 1D). By sampling in a random and systematic manner in 3D at ×630 magnification (using an oil immersion lens), we manually quantified cell density using top of the nucleus as a unique point. Neuron count depended on the presence of Nissl and immunoreactivity for NeuN, while presence of only Nissl allowed us to quantify glial cells. Finally, total cell density was obtained from the sum of neuron and glial cell counts within regions of interest. The Nucleator tool (Gundersen, 1988) was used to quantify neuronal volume using the nucleolus stained with Nissl as a unique point in NeuN⁺ cells and placing a marker. Then, the software computed two perpendicular lines, originating from the defined marker. Finally, the four intersections of the lines with the border of the neuronal cytoplasm based on Nissl staining were marked (Fig. 1D). All measures were quantified throughout the entire regions of interest. The mean number (standard deviation, SD) of NeuN⁺ cells counted per section was 103.813 (30.181), with a mean Schmitz and Hof coefficient of error of 0.102 (0.020) (Schmitz and Hof, 2005).

Quantification of cortical axon density and assessment of axon orientation

Images of axons stained with Bielschowsky silver staining were digitized using a Leica DM/RBE photomicroscope at ×200 magnification attached to a MCID elite image analysis system (Imaging Research Inc, Ontario, Canada). Images were stitched to obtain one image of an entire region of interest that included all cortical layers. Using ImageJ (Schindelin et al., 2012) (version 1.52a, https://imagej.net/Fiji) and an in-house script, axon density was quantified for each region of interest (Fig. 1E). In detail, a segmented line separating the cortex from the white matter and at least three straight lines perpendicular to the cortical surface from the white matter to the pial surface with a distance of >500 µm were manually drawn. The script drew automatically tangent circles with a diameter of 500 µm and with their centres along the straight lines. For each circle, diameter parallel to cortical surface was also drawn. Within all circles defined, the number of axons that crossed the lines perpendicular to the cortex (defined parallel axons), those that crossed the lines parallel to the cortex (defined perpendicular axons), and the total amount of axons (defined as the sum of parallel and perpendicular axons) were counted. When intersecting with both lines, or one line twice, the axon was counted only once according to its preferential direction. To obtain axon density per mm², the total number of axons counted was divided by the total area of circles where the axons were quantified (Fig. 1E).

MRI analysis

White matter T₂-hyperintense lesions were segmented on FLAIR images using multi-view convolutional neural network with batch normalization (Steenwijk et al., 2017) followed by manual editing, yielding lesion maps, which were used to calculate white matter lesion volumes. These volumes were multiplied by the scaling factor obtained from SIENAx software (see below) to correct for head size.

Diffusion tensor MRI processing

Diffusion-weighted images were first corrected for distortions caused by the eddy currents (Haselgrove and Moore, 1996) and the diffusion tensor was estimated by linear regression (Basser et al., 1994). Subsequently, the eigenvalues of the diffusion tensor were calculated by diagonalizing the tensor matrix and maps of fractional anisotropy and mean, axial and radial diffusivities were derived (Pierpaoli and Basser, 1996) (Fig. 2A). All these post-processing steps were performed using FMRIB Software Library (FSL) tools version 5.0.9 (https://fsl.fmrib.ox.ac.uk/fsl/).

Figure 2

Pathology-MRI matching pipeline. Schematic representation of the pipeline applied to diffusion tensor MRI and 3D T₁-weighted sequences processing and subsequent pathologic-MRI matching. See text for further details. L = left; PLP = proteolipid protein; R = right.

3D T₁-weighted processing

White matter lesion maps quantified on FLAIR sequences were registered to the 3D T₁ images, and lesion refilling was performed using LEAP (Chard et al., 2010) to minimize the impact of lesions on subsequent automated segmentations. 3D T₁-weighted images were corrected for signal intensity inhomogeneity using the SPM12 bias field correction and normalized brain volumes, grey matter, cortical grey matter and white matter volumes were assessed using the SIENAx software (Smith et al., 2002). Probability maps of grey matter, white matter and CSF compartments were then created using the standard unified segmentation model in SPM12 (Fig. 2B). Grey matter maps were thresholded at a probability value of 0.5 to limit partial volume effects from white matter and CSF, masked to remove the cerebellum and basal ganglia, and then binarized, to obtain only the cortex.

Histopathology-MRI matching

Brain tissue blocks were carefully matched to the 3D T₁-weighted sequence, with manual identification of MRI regions corresponding to the tissue blocks. This anatomical match was performed by consensus of two experienced observers (P.P. and S.K.). In case of disagreement, a third senior observer (L.J.) made the final decision. To obtain reliable matching, a two-step approach was used. First, we defined the topography of the 10-mm thick coronal brain slices that were cut according to strict standardized protocols (Seewann et al., 2009; Jonkman et al., 2019). Second, we matched the available brain tissue blocks using as many cortical/subcortical anatomical landmarks (e.g. morphology of gyri and sulci, presence of white matter lesions and identification of deep grey matter structures) as possible (Fig. 2C). Matched MRI regions were identified on no more than five subsequent slices on coronal view (maximum total thickness ≤5 mm). Reproducibility of histopathology-MRI matching was evaluated by repeating the manual identification of MRI regions matched with the pathological tissue blocks for 10% (n = 20) of the regions of interest. The median (interquartile range, IQR) dice similarity coefficient was 0.928 (0.910–0.938) (Supplementary Fig. 1).

To limit the analysis on the cortex, regions identified on 3D T₁-weighted sequences were masked with thresholded and binarized grey matter maps. Finally, regions of interest corresponding to normal-appearing and demyelinated cortices were manually defined (Fig. 2C).

Quantification of diffusion tensor MRI metrics in the cortex

After skull stripping by using the Brain Extraction Tool, FSL was used to affinely co-register the 3D T₁-weighted sequences with the distortion-free b = 0 image.

To evaluate distortions possibly associated with EPI sequences, each co-registered 3D T₁-weighted scan with b = 0 image was carefully visually checked.

The calculated transformations were applied to the binary masks of regions of interest identifying normal-appearing and demyelinated cortices (Fig. 2). After transformation, maps of normal-appearing and demyelinated cortices were thresholded at a probability value of 0.5 to reduce partial volume effects and overlaps between different regions.

Finally, the average values of fractional anisotropy and mean, axial and radial diffusivities were calculated in regions of interest of normal-appearing and demyelinated cortices intersecting them with maps of fractional anisotropy and mean, axial and radial diffusivities.

Reproducibility of diffusivity measures was evaluated by re-assessing the diffusion tensor MRI measures for 10% (n = 20) of regions of interest. Intraclass correlation coefficient (95% confidence interval, CI) was 0.991 (0.978–0.996) for fractional anisotropy, 0.991 (0.979–0.997) for mean diffusivity, 0.994 (0.986–0.998) for axial diffusivity and 0.990 (0.975–0.996) for radial diffusivity. Bland-Altman plots are also reported in Supplementary Fig. 1.

Statistical analysis

Demographic variables and post-mortem delay (i.e. time between death and start of autopsy) were compared between non-neurological controls and multiple sclerosis patients using Pearson’s chi-squared or Mann-Whitney tests. Age-, sex- and post-mortem delay-adjusted linear models were applied to assess differences in global MRI metrics. Post-mortem delay was included in the models since MRI measures could be affected by the duration of death-to-scan interval (Miller et al., 2011).

Linear mixed-effects models provided the framework to compare diffusion tensor MRI-derived and pathological measures among non-neurological control cortex, multiple sclerosis normal-appearing cortex and multiple sclerosis demyelinated cortex, accounting for the hierarchical data structure. Random effects were added to model the within-subject correlation among histological samples, even in relation to the tissue type. False discovery rate (FDR) correction was carried out to take the overall number of pairwise contrasts into account.

The potential associations between diffusion tensor MRI measures and pathological substrates in the different tissue types were investigated with similar models, including a Substrate × Tissue type interaction term testing the heterogeneity of the effects. FDR correction was performed to adjust for the overall number of associations tested.

Age, sex and post-mortem delay were included as covariates in all the aforementioned analyses.

A P-value <0.05 was considered statistically significant. SAS (v.9.4) was used for the computation.

Data availability

The data analysed during the current study are available from the corresponding author on reasonable request.

Results

Demographic, clinical and conventional MRI findings

Table 1 and Supplementary Table 1 summarize the main demographic, clinical and conventional MRI findings from non-neurological controls and multiple sclerosis patients.

No significant differences were found for age, sex and post-mortem delay between the two groups (P-values from 0.140 to 0.425) (Table 1). Conversely, compared to non-neurological controls, multiple sclerosis patients showed a higher T₂ lesion volume (P < 0.0001), and cortical mean diffusivity (P = 0.010), axial diffusivity (P = 0.008) and radial diffusivity (P = 0.009), as well as lower normalized brain volume (P = 0.013), grey matter volume (P = 0.032) and white matter volume (P = 0.034) (Table 1).

Distribution of cortical demyelination in patients with multiple sclerosis

One hundred and ten tissue blocks were obtained (56 from non-neurological control subjects and 54 from patients with multiple sclerosis; see Supplementary Table 1 for details). Twenty-seven (50%) of 54 tissue blocks evaluated from multiple sclerosis patients showed only normal-appearing cortex, 5/54 (9.3%) had only multiple cortical lesions, while 22/54 (40.7%) presented both normal-appearing cortex and cortical lesions. No cortical abnormalities were found in non-neurological controls.

Thirty-one cortical lesions were identified from the 54 multiple sclerosis cortical regions. According to their distribution (Bo et al., 2003), cortical lesions were 4/31 (12.9%) type I, 1/31 (3.2%) type II, 24/31 (77.4%) type III and 2/31 (6.5%) type IV.

Histopathological and diffusion tensor MRI findings

Compared to non-neurological control cortex, multiple sclerosis normal-appearing cortex showed lower fractional anisotropy [FDR-corrected P-value (P_FDR) = 0.047], perpendicular axon density (P_FDR = 0.012) and total axon density (P_FDR = 0.028), but not parallel axon density (P_FDR = 0.244) (Table 2 and Fig. 3).

Table 2

Comparisons of diffusion tensor MRI and pathological measures among tissue block non-neurological control cortex, and multiple sclerosis normal-appearing and demyelinated cortices

Variable	Non-neurological control tissue blocks n = 56	Multiple sclerosis patient tissue blocks n = 54		Multiple sclerosis normal-appearing cortex versus non-neurological control cortex		Multiple sclerosis demyelinated cortex versus non-neurological control cortex		Multiple sclerosis demyelinated cortex versus multiple sclerosis normal-appearing cortex
	Cortex	Normal-appearing cortex	Demyelinated cortex
	Mean (SD)	Mean (SD)	Mean (SD)	Estimated mean difference (95% CI)	P	Estimated mean difference (95% CI)	P	Estimated mean difference (95% CI)	P
					PFDR		PFDR		PFDR
FA	0.205 (0.044)	0.174 (0.032)	0.201 (0.043)	−0.028 (−0.048, −0.007)	0.012	0.015 (−0.014, 0.043)	0.277	0.042 (0.021, 0.064))	0.001
					0.047		0.420		0.012
MD, mm2s−1 × 10−3	0.487 (0.087)	0.544 (0.126)	0.528 (0.110)	0.045 (−0.008, 0.097)	0.087	0.018 (−0.056, 0.092)	0.606	−0.027 (−0.091, 0.037)	0.375
					0.191		0.687		0.492
AD, mm2s−1 × 10−3	0.559 (0.095)	0.621 (0.134)	0.613 (0.119)	0.045 (−0.014, 0.105)	0.118	0.024 (−0.059, 0.107)	0.534	−0.021 (−0.090, 0.047)	0.509
					0.246		0.623		0.623
RD, mm2s−1 × 10−3	0.451 (0.084)	0.507 (0.122)	0.484 (0.112)	0.044 (−0.007, 0.095)	0.082	0.006 (−0.069, 0.081)	0.856	0.038 (−0.031, 0.107)	0.254
					0.191		0.856		0.420
Myelin density, % [PLP]	62.251 (12.946)	60.813 (12.274)	43.846 (13.832)	−1.992 (−7.891, 3.906)	0.473	−18.271 (−28.864, −7.678)	0.003	−16.278 (−22.264, −10.293)	<0.0001
					0.602		0.018		0.004
Microglia density, % [Iba-1]	4.008 (1.596)	5.146 (1.949)	5.044 (2.487)	0.653 (−0.611, 1.917)	0.280	1.004 (−0.466, 2.474)	0.161	0.351 (−0.400, 1.103)	0.326
					0.420		0.294		0.472
Astrocyte density, % [GFAP]	11.630 (4.701)	10.254 (4.485)	8.139 (3.943)	−1.928 (−4.659, 0.802)	0.135	−4.091 (−7.366, −0.815)	0.022	−2.163 (−4.314, −0.011)	0.049
					0.257		0.078		0.148
Neuronal density, per mm3 (×104) [NeuN+]	2.950 (0.694)	3.441 (1.051)	3.320 (0.916)	0.305 (−0.252, 0.861)	0.247	0.147 (−0.366, 0.660)	0.533	−0.158 (−0.522, 0.206)	0.352
					0.420		0.623		0.479
Neuronal volume, µm3 (×103) [NeuN+]	2.829 (0.825)	2.782 (0.721)	2.752 (0.800)	−0.029 (−0.383, 0.324)	0.855	−0.095 (−0.620, 0.431)	0.693	−0.065 (−0.428, 0.297)	0.693
					0.856		0.746		0.746
Glial cell density, per mm3 (×104) [Nissl+;NeuN-]	8.557 (1.609)	8.314 (1.728)	7.797 (1.992)	−0.513 (−1.699, 0.673)	0.353	−1.046 (−2.252, 0.160)	0.082	−0.533 (−0.999, −0.620)	0.030
					0.479		0.191		0.095
Total cell density, per mm3 (×104) [Nissl+;NeuN+]	11.507 (1.939)	11.755 (2.529)	11.116 (2.559)	−0.234 (−1.752, 1.284)	0.735	−0.839 (−2.427, 0.750)	0.263	−0.605 (−1.308, 0.099)	0.084
					0.772		0.420		0.191
Perpendicular axon density, per mm2 (×102) [Bielschowsky]	4.465 (0.591)	3.909 (0.612)	3.735 (0.761)	−0.628 (−0.951, −0.305)	0.001	−0.813 (−1.205, −0.421)	0.001	−0.185 (−0.383, 0.013)	0.065
					0.012		0.012		0.181
Parallel axon density, per mm2 (×102) [Bielschowsky]	3.820 (0.586)	3.723 (0.551)	3.392 (0.776)	−0.191 (−0.443, 0.061)	0.123	−0.495 (−0.823, −0.167)	0.007	−0.304 (−0.476, −0.131)	0.003
					0.246		0.029		0.018
Total axon density, per mm2 (×102) [Bielschowsky]	8.285 (1.132)	7.632 (1.094)	7.127 (1.501)	−0.823 (−1.348, −0.299)	0.005	−1.296 (−1.973, −0.618)	0.001	−0.472 (−0.783, −0.161)	0.007
					0.028		0.012		0.029

Variables are reported as mean and standard deviations (SD). Note: fractional anistropy is a dimensionless quantity. PFDR = adjustment for the overall number of comparisons performed [n = 42 (3 pairwise contrasts × 14 variables)]. See main text for further details.

AD = axial diffusivity; Adj = adjusted; FA = fractional anisotropy; Iba-1 = ionized calcium-binding adapter molecule 1; MD = mean diffusivity; NeuN = neuronal nuclear antigen; PLP = proteolipid protein; RD = radial diffusivity.

Figure 3

Between-group comparisons of diffusion tensor-derived and histopathological measures. Box-plots showing between-group differences among non-neurological control cortex, and multiple sclerosis patients normal-appearing and demyelinated cortices. (A) fractional anisotropy; (B) myelin density [proteolipid protein (PLP)]; (C) density of axons perpendicular (⊥) to the cortical surface (Bielschowsky silver staining); (D) density of axons parallel (||) to the cortical surface (Bielschowsky silver staining); (E) total axon density (Bielschowsky silver staining). See main text for further details.

Compared to non-neurological control cortex, multiple sclerosis demyelinated cortex showed lower myelin density (P_FDR = 0.018), perpendicular (P_FDR = 0.012), parallel (P_FDR = 0.029) and total axon density (P_FDR = 0.012) (Table 2 and Fig. 3).

Compared to multiple sclerosis normal-appearing cortex, multiple sclerosis demyelinated cortex showed higher fractional anisotropy (P_FDR = 0.012), lower myelin density (P_FDR = 0.004), parallel (P_FDR = 0.018) and total axon density (P_FDR = 0.029) (Table 2 and Fig. 3).

No between-group differences for mean, axial or radial diffusivity as well as for density of microglia, astrocytes, neuron, glial or total cell and neuronal volume were found (Table 2 and Supplementary Fig. 2).

Associations between diffusion tensor MRI measures and histopathological findings

Table 3, Fig. 4 and Supplementary Fig. 3 summarize the associations between fractional anisotropy and pathological substrates in cortex from non-neurological controls, and multiple sclerosis normal-appearing and demyelinated cortices.

Table 3

Fractional anisotropy substrates in the cortex

Variable	Tissue type	β (95% CI) (×10−2)	P	PFDR	Among-group heterogeneity P	P versus non-neurological control cortex	P versus multiple sclerosis normal- appearing cortex
Myelin density, % (×10) [PLP]	Non-neurological control cortex	0.580 (−0.385, 1.544)	0.203	0.325	0.819	–	0.906
	Multiple sclerosis normal-appearing cortex	0.522 (0.059, 0.986)	0.032	0.106		0.906	–
	Multiple sclerosis demyelinated cortex	0.262 (−0.732, 1.255)	0.561	0.731		0.613	0.556
Microglia density, % [Iba-1]	Non-neurological control cortex	−0.329 (−1.396, 0.739)	0.490	0.669	0.402	–	0.960
	Multiple sclerosis normal-appearing cortex	−0.303 (−0.733, 0.127)	0.139	0.261		0.960	–
	Multiple sclerosis demyelinated cortex	−0.722 (−1.427, −0.0162)	0.046	0.122		0.505	0.192
Astrocyte density, % [GFAP]	Non-neurological control cortex	−0.182 (−0.413, 0.049)	0.099	0.198	0.624	–	0.377
	Multiple sclerosis normal-appearing cortex	−0.054 (−0.312, 0.203)	0.610	0.756		0.377	–
	Multiple sclerosis demyelinated cortex	−0.075 (−0.904, 0.753)	0.825	0.853		0.761	0.943
Neuronal density, per mm3 (×104) [NeuN+]	Non-neurological control cortex	−1.200 (−3.202, 0.803)	0.200	0.325	0.069	–	0.090
	Multiple sclerosis normal-appearing cortex	0.586 (−0.435, 1.606)	0.217	0.325		0.090	–
	Multiple sclerosis demyelinated cortex	1.588 (0.234, 2.943)	0.028	0.106		0.026	0.183
Neuronal volume, µm3 (×103) [NeuN+]	Non-neurological control cortex	−0.137 (−1.123, 0.849)	0.752	0.806	0.423	–	0.822
	Multiple sclerosis normal-appearing cortex	−1.062 (−2.361, 0.236)	0.094	0.198		0.206	–
	Multiple sclerosis demyelinated cortex	0.038 (−1.614, 1.689)	0.958	0.958		0.822	0.339
Glial cell density, per mm3 (×104) [Nissl+;NeuN-]	Non-neurological control cortex	−0.179 (−0.737, 0.378)	0.472	0.669	0.092	–	0.041
	Multiple sclerosis normal-appearing cortex	0.733 (0.091, 1.374)	0.031	0.106		0.041	–
	Multiple sclerosis demyelinated cortex	0.662 (−0.074, 1.397)	0.071	0.164		0.073	0.843
Total cell density, per mm3 (×104) [Nissl+;NeuN+]	Non-neurological control cortex	−0.248 (−0.667, 0.172)	0.206	0.325	0.033	–	0.028
	Multiple sclerosis normal-appearing cortex	0.471 (0.003, 0.940)	0.049	0.122		0.028	–
	Multiple sclerosis demyelinated cortex	0.677 (0.121, 1.233)	0.024	0.106		0.015	0.444
Perpendicular axon density, per mm2 (×102) [Bielschowsky]	Non-neurological control cortex	2.494 (1.023, 3.966)	0.005	0.031	0.307	–	0.293
	Multiple sclerosis normal-appearing cortex	1.705 (0.874, 2.536)	0.002	0.031		0.293	–
	Multiple sclerosis demyelinated cortex	0.445 (−2.112, 3.002)	0.693	0.77		0.150	0.302
Parallel axon density, per mm2 (×102) [Bielschowsky]	Non-neurological control cortex	3.288 (1.408, 5.168)	0.004	0.031	0.114	–	0.169
	Multiple sclerosis normal-appearing cortex	1.662 (0.014, 3.310)	0.049	0.122		0.169	–
	Multiple sclerosis demyelinated cortex	0.466 (−1.897, 2.829)	0.655	0.756		0.064	0.445
Total axon density, per mm2 (×102) [Bielschowsky]	Non-neurological control cortex	1.561 (0.686, 2.436)	0.003	0.031	0.170	–	0.184
	Multiple sclerosis normal-appearing cortex	0.934 (0.448, 1.421)	0.002	0.031		0.184	–
	Multiple sclerosis demyelinated cortex	0.259 (−0.983, 1.500)	0.644	0.756		0.085	0.323

Associations between fractional anisotropy from diffusion tensor MRI measures and pathological substrates in non-neurological control cortex, and multiple sclerosis normal-appearing and demyelinated cortices. Substrate × Tissue type interactions to evaluate the heterogeneity of the effects are also shown. Note: β are expressed as ×10⁻² for each increase of 10% of myelin density, 1% of microglia density, 10⁴ cells/mm³ or 10² axons/mm², respectively.

pFDR = adjustment for the overall number of associations tested [n = 30 (3 types of tissue × 10 pathological variables)].

See main text for further details.

Iba-1 = ionized calcium-binding adapter molecule 1; NeuN = neuronal nuclear antigen; PLP = proteolipid protein. Statistically significant comparisons (P < 0.05 or P_FDR < 0.05) are shown in bold.

Figure 4

Fractional anisotropy substrates in the cortex. Scatterplots showing significant associations between fractional anisotropy derived from diffusion-tensor MRI measures and pathological substrates in non-neurological control cortex (blue), multiple sclerosis normal-appearing cortex (yellow), and multiple sclerosis demyelinated cortex (red). (A) density of axons perpendicular to the cortical surface (Bielschowsky silver staining); (B) density of axons parallel to the cortical surface (Bielschowsky silver staining); (C) total axon density (Bielschowsky silver staining). Continuous line = statistically significant associations; dotted line = not statistically significant associations. (D) Heat map showing betas (βs) of associations between fractional anisotropy and pathological substrates. Statistically significant betas are shown in bold. See text for further details.

In cortex from non-neurological controls, fractional anisotropy was positively associated with perpendicular (P_FDR = 0.031), parallel (P_FDR = 0.031) and total axon density (P_FDR = 0.031) (Table 3 and Fig. 4).

In multiple sclerosis normal-appearing cortex, fractional anisotropy was positively associated with perpendicular (P_FDR = 0.031), and total axon density (P_FDR = 0.002) (Table 3 and Fig. 4). Positive associations were also found between fractional anisotropy and myelin density (P = 0.032), glial cell density (P = 0.031), total cell density (P = 0.049), and parallel axon density (P = 0.049), but none survived correction for multiple comparisons (Table 3 and Supplementary Fig. 3).

In multiple sclerosis demyelinated cortex, fractional anisotropy was negatively associated with microglia density (P = 0.046) and positively associated with neuron density (P = 0.028) and total cell density (P = 0.024), but none survived correction for multiple comparisons (Table 3 and Supplementary Fig. 3).

Significant Substrate × Tissue type interactions were found in non-neurological control cortex versus multiple sclerosis normal-appearing cortex for glial cell density (P = 0.041) and total cell density (P = 0.028) and in non-neurological control cortex versus multiple sclerosis demyelinated cortex for neuron density (P = 0.026) and total cell density (P = 0.015) (Table 3).

Discussion

By combining post-mortem in situ diffusion tensor MRI and a wide array of histopathological markers, this study crucially identified the pathological substrates of diffusion tensor MRI changes in multiple sclerosis normal-appearing and demyelinated cortices. In non-neurological control cortex, fractional anisotropy values are associated with axons perpendicular to cortical surface that represent the principal cortical efferent projections (Fig. 5). In multiple sclerosis normal-appearing cortex, a decrease of fractional anisotropy and a loss of these projective axons occur compared to non-neurological control cortex (Fig. 5). The concomitant loss of perpendicular and parallel axons could increase coherence of multiple sclerosis demyelinated cortex, thus increasing fractional anisotropy (Fig. 5).

Figure 5

Proposed explanation of fractional anisotropy patterns in the cortex. Top row: Schematic representations of cortical neurons and parallel and perpendicular axons in (A) non-neurological control cortex, (B) multiple sclerosis normal-appearing cortex, and (C) multiple sclerosis demyelinated cortex. Bottom row: Corresponding diffusion ellipsoids. In non-neurological control cortex (A), fractional anisotropy values are suggested to be mainly driven by perpendicular axons, that represent the principal cortical efferent projections and are characterized by an orientation orthogonal to cortical surface. In multiple sclerosis normal-appearing cortex (B), the significant decreased fractional anisotropy found compared to non-neurological control cortex could be explained by a secondary retrograde degeneration of these projective axons due to focal demyelinating lesions and diffuse abnormalities occurring in the white matter. In multiple sclerosis demyelinated cortex (C), although a significant loss of perpendicular axons occurred, the concomitant reduction of parallel axons could increase the coherence of the tissue containing cortical lesions (in pink), thus determining an increase of fractional anisotropy.

Consistent with previous literature (Geurts and Barkhof, 2008; Calabrese et al., 2010a; Magliozzi et al., 2018; Scalfari et al., 2018; Filippi et al., 2019), we found that the cortex was frequently affected by demyelination. Cortical lesions were present in 27/54 (50%) multiple sclerosis patient tissue blocks and the majority of them (77.4%) were type III lesions (Bo et al., 2003), thus confirming the high prevalence of subpial demyelination in progressive multiple sclerosis.

By evaluating diffusion tensor MRI measures commonly investigated as indicators of tissue pathology (Filippi et al., 2017), our post-mortem in situ MRI assessment showed that a significant decrease of fractional anisotropy occurred in multiple sclerosis normal-appearing cortex compared to non-neurological control cortex. Similar to previous studies (Poonawalla et al., 2008; Calabrese et al., 2011; Filippi et al., 2013; Jonkman et al., 2016b; Yaldizli et al., 2016; Preziosa et al., 2017), fractional anisotropy in multiple sclerosis demyelinated cortex was significantly increased compared to multiple sclerosis normal-appearing cortex, and similar to non-neurological control cortex. Conversely, mean, axial and radial diffusivities did not significantly differ among tissue types. This latter finding could be due to the high variability of diffusivity values found in both multiple sclerosis normal-appearing and demyelinated cortices, reflecting heterogeneous pathological mechanisms (e.g. demyelination, neuro-axonal loss and gliosis), but also body temperature, acting in opposite directions on water molecule diffusion.

Different pathological substrates have been hypothesized to explain such fractional anisotropy changes. Since demyelination is the pathological hallmark of cortical lesions (Geurts and Barkhof, 2008; Calabrese et al., 2010a), we evaluated the relations between cortical myelin density and fractional anisotropy values first. Although myelin density and microarchitecture are different in the cortex compared to white matter, and since a strong positive association between myelin content and fractional anisotropy occurs in the white matter (Schmierer et al., 2007), myelin should contribute (albeit partially) to cortical diffusivity properties. In our study, demyelinated cortex showed a significant myelin loss compared to both non-neurological control cortex and multiple sclerosis normal-appearing cortex, but no significant associations were found between myelin density and fractional anisotropy. Although histological heterogeneities among the cortical regions investigated could contribute to this lack of findings, the concomitant presence of a significantly higher fractional anisotropy and a lower myelin density in demyelinated cortex suggests that other pathological substrates are likely to explain such an increased fractional anisotropy.

A significantly higher amount of activated microglia has been suggested to explain diffusion tensor MRI abnormalities (Poonawalla et al., 2008; Calabrese et al., 2011; Filippi et al., 2013; Preziosa et al., 2017). An increased size of microglial perikarya, thicker processes and an orientation more perpendicular to the pial surface characterize activated microglia (Peterson et al., 2001) and might contribute to increase tissue anisotropy. Although a significant amount of activated microglia can be detected in cortical biopsy from early, atypical multiple sclerosis cases (Lucchinetti et al., 2011), and in a subset of multiple sclerosis patients, with a shorter disease duration and a pro-inflammatory profile (Kooi et al., 2012), cortical lesions typically have a low degree of activated microglia (Peterson et al., 2001; Calabrese et al., 2010a). By including long-standing progressive multiple sclerosis patients (median disease duration = 31.4 years), we found that microglia, quantified using a general microglial marker (Iba-1), was not significantly different between non-neurological controls and multiple sclerosis patients, also independently from cortical demyelination. Although activated microglia was not directly evaluated, our results are in line with a recent 7 T post-mortem MRI/histopathology study (Jonkman et al., 2016b) showing an increased fractional anisotropy in cortical lesions, but only a few activated microglia (quantified using LN3) in the demyelinated cortex. All these findings suggest that microglia does not strongly contribute to fractional anisotropy changes.

Astrocytes are among the most common cells in the brain, characterized by a wide range of morphologies, and involved in structural and metabolic support, and maintenance of homeostasis (Khakh and Sofroniew, 2015; Ludwin et al., 2016). While reactive astrocytosis is common in multiple sclerosis, mainly in active white matter plaques (Ludwin et al., 2016), this reactivity is lower in cortical lesions and is absent in inactive plaques where a limited number of small astrocytes is visible (Ludwin et al., 2016). By evaluating cortical lesions in patients with long-standing multiple sclerosis, our study found that the density of astrocytes, quantified using GFAP, was not significantly different between non-neurological control subjects and patients with multiple sclerosis, also independently from cortical demyelination. Moreover, astrocyte density was not significantly associated with fractional anisotropy values, suggesting that astrocyte density changes do not strongly contribute to cortical fractional anisotropy differences.

A variable degree of neurodegenerative phenomena involving neurons and glial cells have been demonstrated in both normal-appearing and demyelinated cortices (Peterson et al., 2001; Wegner et al., 2006; Magliozzi et al., 2007, 2010; Vercellino et al., 2007; Klaver et al., 2015; Jurgens et al., 2016; Carassiti et al., 2017). Neuronal shrinkage and loss of neurons, axons and synapses (Peterson et al., 2001; Wegner et al., 2006; Vercellino et al., 2007; Magliozzi et al., 2010; Klaver et al., 2015; Jurgens et al., 2016; Carassiti et al., 2017) could determine microstructural tissue abnormalities that influence diffusion tensor MRI measures. We found no significant differences in neuronal or glial density and volume between non-neurological controls and multiple sclerosis patients, independently from cortical demyelination. This suggests a limited role of cell density and volume in explaining cortical fractional anisotropy heterogeneities. Nevertheless, this lack of findings seems in contrast to previous studies (Peterson et al., 2001; Wegner et al., 2006; Vercellino et al., 2007; Magliozzi et al., 2010; Klaver et al., 2015; Carassiti et al., 2017). Heterogeneous characteristics of multiple sclerosis donors and of cortical brain areas evaluated (Magliozzi et al., 2010; Klaver et al., 2015) but also differences in methods applied to stain neurons (NeuN, Giemsa, or anti-neural GAP-43) or to investigate neuronal properties (stereology or 2D morphometric analyses) could explain these discrepancies. Moreover, differently from other studies (Wegner et al., 2006; Magliozzi et al., 2010), we quantified neuronal density and volume considering all types of neurons in all cortical layers. This was because all cells with different neuronal shapes, sizes and densities (Nieuwenhuys, 1994) could affect diffusivity properties and our MRI sequences did not have enough spatial resolution to discriminate cortical layers. Of note, a recent 7 T post-mortem MRI/histopathology study showed that cortical lesions, compared to normal-appearing cortex, were characterized by a higher cell density, possibly as consequence of tissue compaction (Jonkman et al., 2016b). Neuro-axonal damage, combined with loss of myelin and dendritic arborization, could cause tissue compaction in multiple sclerosis (Popescu et al., 2015; Jonkman et al., 2016b). This could determine a paradoxical increase of neuronal and glial cell density and counteract the multiple sclerosis-related neuronal loss. Taking this into account, and considering the lack of significant association between neuronal density and fractional anisotropy values except for demyelinated cortex, our study suggest that neuronal volume and density does not strongly contribute to fractional anisotropy.

Axonal density and orientations strongly contribute to white matter diffusion tensor MRI properties (Schmierer et al., 2007). In line with previous studies (Peterson et al., 2001; Magliozzi et al., 2007; Klaver et al., 2015; Jurgens et al., 2016), we found that a significant axonal loss occurred in patients with multiple sclerosis, being more severe in demyelinated cortex. Interestingly, in multiple sclerosis, compared to non-neurological controls, a significant loss of perpendicular axons was found, with no significant difference between normal-appearing and demyelinated cortices. These perpendicular axons represent the principal cortical efferent projections to basal ganglia, brainstem and spinal cord (Nieuwenhuys, 1994; Migliore and Shepherd, 2005; Molyneaux et al., 2007). Because of their primarily perpendicular distribution, they promote an orientation orthogonal to cortical surface of the major diffusion tensor eigenvector and thus higher fractional anisotropy values (McKinstry et al., 2002; Miller et al., 2011; Ouyang et al., 2019) (Fig. 5). This has been shown in post-mortem studies (Miller et al., 2011; Aggarwal et al., 2015), but also in studies investigating brain development (McKinstry et al., 2002; Ouyang et al., 2019). In particular, at the 26th week of gestational age, when the human foetal cortex is characterized by high fractional anisotropy values (McKinstry et al., 2002; Ouyang et al., 2019), radial neuronal migration from the ventricular zone to the cortical plate is almost complete, but cortical maturation of dendritic harborizations, synapses and afferent fibres has still to begin, determining a highly anisotropic tissue with the main eigenvector perpendicular to the cortical surface. Similar findings have been also demonstrated during brain maturation in other mammals, including mice (Verma et al., 2005) and baboons (Kroenke et al., 2005). In line with this, we found significant positive associations between fractional anisotropy values and perpendicular axon densities in non-neurological control cortex and in multiple sclerosis normal-appearing cortex. The significant decreased fractional anisotropy found in multiple sclerosis normal-appearing cortex could be due, at least partially, to the loss of these projective axons that are highly coherent, oriented perpendicular to the cortical surface and with large length and size (Fig. 5). Focal demyelinating lesions and diffuse abnormalities of white matter tracts comprising these axons could contribute to their secondary retrograde degeneration, independently from local cortical pathology (Preziosa et al., 2011; Haider et al., 2016) (Fig. 5).

Conversely, similar to a recent study (Jurgens et al., 2016), we found a significant reduction of cortical parallel axon density only in multiple sclerosis demyelinated cortex compared to both non-neurological control and multiple sclerosis normal-appearing cortices. These tangentially-oriented cortical fibres mostly represent collateral and terminal branches of afferent intra- and extra-cortical axons, but also axons of interneurons and collateral branches of pyramidal cells (Nieuwenhuys, 1994). Cortical lesions could promote their neurodegeneration because of their particular sensitivity to inflammation, demyelination and oxidative stress (Friese et al., 2014) (Fig. 5). According to this, although a significant loss of perpendicular axons occurred also in demyelinated cortex, the concomitant reduction of parallel axonal density could increase the coherence of the tissue containing cortical lesions, determining an increase of fractional anisotropy (Fig. 5). Supporting evidence to this hypothesis come again from studies applying diffusion tensor MRI to study brain development in humans (McKinstry et al., 2002; Ouyang et al., 2019) and other mammals (Kroenke et al., 2005; Verma et al., 2005). In humans, from the 36th week of gestational age, the sprouting of basal dendrites of pyramidal cells, the formation of local intra-cortical circuits, and the growth of afferent fibres restrict water displacement more uniformly in all directions. This reduces the strength of the principal eigenvector and of fractional anisotropy compared to previous phases of cortical maturation (Nieuwenhuys, 1994; Migliore and Shepherd, 2005; Molyneaux et al., 2007). In multiple sclerosis demyelinated cortex, both parallel and perpendicular axon densities are significantly reduced, resulting both in a ratio between perpendicular and parallel fibres more similar to non-neurological control cortex and a higher fractional anisotropy. Our study also found that fractional anisotropy values were significantly positively associated also with the density of parallel axons in non-neurological control cortex, which could be counterintuitive since more parallel axons would contribute to decrease fractional anisotropy. The strong correlations between perpendicular and parallel axonal densities (r-values from 0.827 to 0.902) could contribute to explain also this association. Moreover, projective fibres have generally larger size and length compared to parallel fibres (Rockland, 2017; Edwards et al., 2018), thus their contribution to diffusivity properties should be larger. Finally, a recent study (Aggarwal et al., 2015) showed a lower cortical fractional anisotropy in regions with a high density of parallel fibres (layer I, inner and outer inner bands of Baillarger), while fractional anisotropy is higher where fibres have a predominant vertical orientation.

Our study has some limitations. First, MRI scans were acquired with pixel sizes optimized for each sequence; therefore, the images were resampled to diffusion image resolution. Moreover, our histopathological analyses were performed across all cortical layers, since sequence resolution did not allow their parcellation. Second, even though post-mortem MRI sequences were acquired with the brain in situ, thus in a setting closer to in vivo condition than ex vivo (fixated) brain sample scanning, several factors (i.e. changes in body temperature, variable post-mortem delay until scanning, metabolic changes, swelling, hypoxia) influence diffusivity properties (Miller et al., 2011). This should be kept in mind in the comparisons between post-mortem and in vivo studies (Miller et al., 2011). Third, we studied only progressive multiple sclerosis patients with a long-standing disease. Different pathological substrates could predominantly influence diffusion tensor MRI measures in the earlier phases of the disease. For instance, a more relevant inflammatory activity could characterize earliest phases of multiple sclerosis (Granberg et al., 2017). Fourth, we only investigated diffusion tensor MRI to evaluate cortical microstructural tissue abnormalities. More advanced diffusion MRI models, such as neurite orientation dispersion and density imaging (Granberg et al., 2017) and other quantitative MRI techniques, such as magnetization transfer MRI, T₁- and T₂-relaxometry, R2*, and T₁/T₂ ratio (Gracien et al., 2016a, b; Jonkman et al., 2016a; Nakamura et al., 2017; Righart et al., 2017) could be applied to confirm our findings and to investigate more accurately the different pathological substrates affecting multiple sclerosis cortex sclerosis.

In conclusion, demyelination and loss of axons with different orientation ratios, occur in both multiple sclerosis normal-appearing and demyelinated cortices. Reduction of perpendicular axons in multiple sclerosis normal-appearing cortex and of both perpendicular and parallel axons in demyelinated cortex represent the main pathological substrates determining fractional anisotropy changes occurring in multiple sclerosis normal-appearing and demyelinated cortices. Our results suggested that diffusion tensor MRI could contribute to disentangle the pathological substrates of cortical damage in multiple sclerosis that represent a relevant contributor to disease-related disability progression.

Acknowledgements

The authors would like to thank the donors, the Netherlands brain bank, the Normal Aging Brain Collection Amsterdam and autopsy teams. Additionally, Ms Eliane Kaaij for managing NABCA registrations, Prof. Dr Barkhof for performing post-mortem MRI reports, and technicians at the department of Anatomy and Neuroscience for their help in processing the data.

Funding

S.K. received Research support from the Dutch MS Research Foundation, grant number 14–358e.

Competing interests

P.P. received speakers honoraria from Biogen Idec, Novartis, Merck Serono and Excemed. S.K. received research support from the Dutch MS Research Foundation, grant number MS14–358e. W.D.J.vdB. was financially supported by a grant from Amsterdam Neuroscience, ZonMW Memorabel, ZonMW Technology Hotel, Stichting Parkinson Fonds, Alzheimer Netherlands-LECMA, Roche Pharma, Lysosomal Therapeutics, Crossbeta Sciences. She is a consultant for CHDR Leiden and Lysosomal Therapeutics. M.A.R. received speakers honoraria from Biogen Idec, Novartis, Genzyme, Sanofi-Aventis, Teva, Merck Serono and Roche and receives research support from the Italian Ministry of Health and Fondazione Italiana Sclerosi Multipla. M.F. is Editor-in-Chief of the Journal of Neurology; has received compensation for consulting services and/or speaking activities from Biogen Idec, Merck-Serono, Novartis, and Teva Pharmaceutical Industries; and receives research support from Biogen Idec, Teva Pharmaceutical Industries, Novartis, Roche, Italian Ministry of Health, Fondazione Italiana Sclerosi Multipla, and ARiSLA (Fondazione Italiana di Ricerca per la SLA). J.J.G.G. is an editor of MS Journal. He serves on the editorial boards of Neurology and Frontiers of Neurology and is president of the Netherlands organization for health research and innovation. He has served as a consultant for Merck-Serono, Biogen, Novartis, Genzyme and Teva Pharmaceuticals. M.D.S., A.G.M.M., G.J.S. and L.E.J. have nothing to disclose.

Author notes

Paolo Preziosa and Svenja Kiljan authors contributed equally to this work.