Diffusion kurtosis imaging allows the early detection and longitudinal follow-up of amyloid-β-induced pathology

In this study, male WT C57BL/6 J mice (n = 52) and male transgenic APP_KM670/671NL/PS1_L166P mice were used (n = 67, referred to as APP/PS1 mice) [11]. Mice were housed in the animal facility of the University of Antwerp during the whole experiment. During the study, mice were kept on a normal 12-h/12-h day-night cycle with ad libitum access to food and water. Additional file 1 shows the weight evolution of the mice in the longitudinal cohort.

APP/PS1 mice start developing Aβ plaques from the age of 6 to 8 weeks and show aggressive amyloidosis in subsequent months [11]. As such, we acquired the first dataset when mice were 2 months old, when only low Aβ plaque deposition is present (n = 20 WT mice and n = 19 APP/PS1 mice). We then longitudinally followed these mice by acquiring a DKI datasets when they were 4 and 6 months of age (intermediate Aβ plaque load) and finally at 8 months of age, which corresponds to an extensive Aβ plaque load. After acquisition of this last DKI dataset at 8 months of age, mice were then sacrificed for histological analysis. In addition, three further cohorts of WT and APP/PS1 mice were scanned once each at 2 months (n = 11 WT mice and n = 16 APP/PS1 mice), 4 months (n = 10 WT mice and n = 16 APP/PS1 mice) and 6 months (n = 11 WT mice and n = 16 APP/PS1 mice) of age and killed thereafter for histological analysis. The complete experimental design is shown in Fig. 1.

At 2, 4, 6 and 8 months of age, mice were subjected to ¹H-MRI scanning using a 7-T PharmaScan MRI scanner with a 16-cm diameter horizontal bore (Bruker, Bremen, Germany). This system is equipped with a standard Bruker cross-coil setup using a quadrature volume coil for excitation and an array mouse surface coil for signal detection. The system was interfaced with a Linux PC running TopSpin 2.0 and Paravision 5.1 software (Bruker BioSpin, Ettlingen, Germany). Anaesthesia was induced using 2% isoflurane (Abbott, Maidenhead, UK) in a gas mixture of 30% O₂ and 70% N₂ at a flow rate of 600 ml/minute. During MR image acquisition, the isoflurane concentration was initially set at 2% and subsequently lowered when required to maintain a stable respiration rate of (110 ± 10) breaths per minute, which was monitored using a pressure-sensitive pad. In addition, body temperature was monitored via a rectal probe and was held constant between 37.0 °C and 37.3 °C using warm air coupled to a feedback unit (SA Instruments, Stony Brook, NY, USA). PC-sam monitoring software (SA Instruments) was used to measure respiration rate and to measure and control body temperature. Following MR image acquisition, mice were left to recover separately under a heating lamp before being returned to their respective cages.

To ensure uniform slice positioning, we first acquired axial, sagittal and horizontal multi-slice 2D rapid acquisition and relaxation enhancement (RARE) images using the following parameters: repetition time (TR) = 2500 ms, echo time (TE) = 33 ms, matrix size (256 × 256), field of view (FOV) = (20 × 20) mm², resolution = (0.078 × 0.078) mm², nine slices, slice thickness = 0.8 mm, and RARE factor = 8. Following correct positioning, we acquired three DKI scans, each with 20 unique diffusion gradient directions and seven b values (400, 800, 1200, 1600, 2000, 2400 and 2800 s/mm²). In addition, seven images without diffusion weighting (b₀) were included for each DKI scan. This yielded DKI data comprising 21 b₀ images and diffusion weighting for 60 diffusion directions with seven b values. Images were collected with a multi-slice two-shot spin-echo/echo planar imaging sequence with the following parameters: TR = 7000 ms, TE = 23.25 ms, δ = 4 ms, Δ = 12 ms, acquisition matrix = (96 × 96), spatial resolution = (0.214 × 0.214) μm², 28 horizontal slices, and slice thickness = 0.20 mm. This resulted in a total acquisition time of 1 h, 43 minutes. Next, a high-resolution 3D anatomical image was acquired using a T2-weighted 3D RARE sequence in the same horizontal orientation as the DKI data. The following parameters were used: TR = 3185 ms, TE = 44 ms, and spectral width = 50 kHz, averages = 1, RARE factor = 8, matrix size = (265 × 64 × 50), FOV (20.5 × 13.0 × 10.0) mm³, resolution = (0.080 × 0.203 × 0.200) mm³, and total acquisition time = 21 minutes.

Prior to the actual image analysis, we performed a visual inspection for quality of the acquired raw data (e.g., ghosting and/or movement) combined with a semi-automated quality control to avoid bias in the diffusion parameter estimation due to MR acquisition artefacts. This semi-automated data quality control consisted of (1) validating the signal decay with increasing b values, (2) validating if signals obtained with high b values did not reach the noise level, (3) validating the magnitude of the signal-to-noise ratio, and (4) validating the parametric estimation error using a chi-square test (see below). Once quality was assured, image pre-processing and analysis were initiated. The data were corrected for motion and eddy current artefacts using the Functional Magnetic Resonance Imaging of the Brain (FMRIB) Software Library (FSL) [13].

The DKI model is shown below:with S(0) being the signal intensity without diffusion weighting, D being the rank 2 diffusion tensor, W being the rank 4 kurtosis tensor, b being the b value, and g being the diffusion gradient direction. The DKI model was voxel-wise fitted to the diffusion-weighted images. The diffusion tensor and kurtosis tensor quantify the Gaussian diffusion profile and the deviation from a Gaussian diffusion distribution, respectively [10, 14]. The diffusion tensor and the diffusion kurtosis tensor were estimated simultaneously using conditional least squares estimators while imposing positivity on the kurtosis coefficients [15]. The conditional least squares estimator explicitly accounts for the Rician MR data distribution, for which the noise level has been estimated [16]. Rotational invariant parameter maps of axial diffusivity, radial diffusivity (RD), mean diffusivity (MD) and fractional anisotropy (FA) were computed from the diffusion tensor, and further refered to as diffusion tensor (DT) metrics, and axial kurtosis (AK), radial kurtosis (RK) and mean kurtosis (MK) maps were computed from the diffusion kurtosis tensor, and further refered to as diffusion kurtosis (DK) metrics, using MATLAB software (Mathworks Inc., Natick, MA, USA) [14, 17].

A study-based atlas was constructed with Advanced Normalization Tools [18] software using 25 randomly selected 3D T2-weighted MRI datasets across both genotypes and all ages. We delineated 23 grey matter and white matter regions of interest (ROIs) on this atlas using AMIRA (version 5.4) (Additional file 2). All individual 3D T2-weighted MRI scans were then normalised to the atlas. The inverse transformation of the normalisation was used to map the atlas ROIs onto the native space of the individual 3D T2-weighted MRI scans. Afterwards, all individual b₀ MR images were co-registered to their corresponding 3D T2-weighted MRI scans using FSL software. Using the inverse transformation of the latter, we mapped the individual 3D ROIs to the individual DT and DK metric maps. On the basis of optimal contrast of the FA map to differentiate grey matter and white matter, we finally manually checked and, if necessary, corrected the contours of the grey matter and white matter ROIs to limit a possible partial volume effect. Finally, ROI-averaged DT and DK metrics (MD, axial diffusivity, RD, FA and MK, AK and RK) were extracted from the respective metric maps.

For histological examination, mice were sacrificed by means of cervical dislocation. The complete brain was dissected and fixed in Fade4 fixative. Fixed brains were sent to HistoGeneX (Antwerp, Belgium), where 5-μm-thick sagittal paraffin-embedded sections were cut from the left hemisphere. Sectioning was started at the middle of the brain, and ten sagittal sections were acquired at 150-μm intervals, covering the whole hemisphere. Immunohistochemical (IHC) staining for Aβ plaques was initiated by depigmenting slides using potassium permanganate for 3 minutes, followed by oxalic acid for 1 minute. Slides were pre-treated in formic acid for 10 minutes to retrieve epitopes and were then incubated for 15 minutes at room temperature with a mouse anti-Aβ (clone 4G8) antibody (1:20,000, SIG-39200; Eurogentec, Angers, France). Next, slides were incubated with a labelled polymer (Dako EnVision + System-HRP Labeled Polymer Anti-Mouse, K4001; Dako). Finally, the substrate was visualised using 3,3′-diaminobenzidine (DAB) chromogen (Dako Liquid DAB+ substrate chromogen system; Dako) for 5 minutes. All steps were performed using the automated Lab Vision Autostainer 480S (Thermo Scientific, Waltham, MA, USA). In every 4G8 staining run, an immunoglobulin G (IgG) control (mouse IgG2b; Dako) was also included. For myelin basic protein (MBP) staining, epitope retrieval was performed in Target Retrieval Solution (Dako) for 30 minutes at 97 °C. After endogenous peroxidase activity was quenched, the slides were incubated for 30 minutes at room temperature with a mouse anti-MBP (clone SMI-94) antibody (1:5000, SMI-94R; Covance Antibody Products, Princeton, NJ, USA). Next, slides were incubated with a labelled polymer (Dako EnVision + System-HRP Labeled Polymer Anti-Mouse, K4001). Finally, the substrate was visualised using DAB chromogen for 5 minutes. All steps were performed on the automated Lab Vision Autostainer 480S. In every MBP staining run, an IgG control (mouse IgG1; Abcam, Cambridge, UK) was included. For microgliosis (ionised calcium-binding adapter molecule 1 [IBA1]) IHC staining, epitope retrieval was performed in citrate buffer (pH 6; Lab Vision) for 30 minutes at 97 °C. After quenching endogenous peroxidase activity, the slides were incubated for 30 minutes at room temperature with a rabbit anti-IBA1 antibody (1:5000, 019-19741; Wako Pure Chemical Industries, Osaka, Japan). Next, slides were incubated with a labelled polymer (Dako EnVision + System-HRP Labelled Polymer Anti-Rabbit, K4003). Finally, the substrate was visualised using the DAB chromogen (Dako Liquid DAB+ substrate chromogen system) for 5 minutes. All steps were performed using the automated Lab Vision Autostainer 480S. In every IBA1 staining run, an IgG control (rabbit IgG; Dako) was included. For astrogliosis (glial fibrillary acidic protein [GFAP]) IHC staining, epitope retrieval was performed using cell conditioning solution (Ventana Medical Systems, Tucson, AZ, USA). The slides were incubated for 28 minutes at 37 °C with a rabbit anti-GFAP antibody (1:7500, Z0334; Dako). Next, slides were incubated with an OmniMap anti-rabbit HRP detection system (Ventana Medical Systems). All steps were performed on the automated Ventana Discovery® XT platform (Ventana Medical Systems). In every GFAP staining run, an IgG control (rabbit IgG; Dako) was also used. For all four staining runs, stained slides were scanned using the MIRAX digital slide scanner (Carl Zeiss Microscopy, Göttingen, Germany).

Quantification of the stained slides was performed by DCILabs (Keerbergen, Belgium). In short, the histological images are colour de-convolved to separate the colours on the stained slides [19]. This generates a grey value image of the MBP-, GFAP-, 4G8- and IBA1-stained slides. Hysteresis thresholds were applied on these images to determine the percentage of area positive for the respective staining (percent optical density [%O.D.]) and for each of these ROIs. To quantify the amount of 4G8-positive Aβ plaques, a mask based on an intensity cutoff value to distinguish Aβ plaques from background signal in the grey value images was compiled, which then automatically counted the amount of objects in the mask. To assess the presence of elongated structures in the image, an anisotropic measure was defined as the ratio of the difference and the sum of the eigenvectors of the local structure tensor [20]. All analyses were performed with software developed in-house using OpenSlide C interface (openslide.org) to read the MIRAX images.

Co-registration of all images of the histological staining was performed, and an overview of the whole co-registration pipeline is shown in Fig. 2. In short, because the MBP staining contains considerable anatomical information, we used the MBP-stained slide of each sectioned interval (150 μm apart) to create a 3D histological reference space. These MBP-stained slides were stacked onto a 3D dataset using an algorithm previously described by Lowe et al. [21] and Ourselin et al. [22]. Next, we co-registered this 3D histological stack to the 3D T2-weighted MRI atlas as described by Ourselin et al. [22] and Modat et al. [23]. Next, a moment-matching algorithm was used to match the 4G8, GFAP and IBA1 staining with the MBP images in order to propagate all histological stains to the corresponding MR images. We then combined all MRI and histological data and created a relational database which contained the following per voxel: the MRI atlas coordinates, the ROI information, all diffusion and diffusion kurtosis metrics, and the corresponding histological values. On the basis of this database, we performed the statistical analyses described below.

We have previously shown that DKI is able to visualise Aβ plaque-induced pathology in APP/PS1 mice at 16 months of age [12]. In the present study, we first investigated the ability of DKI to visualise early and progressive Aβ plaque-induced pathology in APP/PS1 mice at 2, 4, 6 and 8 months of age. As such, we performed an ROI-based analysis of the longitudinally acquired dataset. The most meaningful statistically significant differences for different metrics of different ROIs are shown in Fig. 3.

In the motor cortex, we observed a genotype effect for axial diffusivity and an age effect for axial diffusivity and RD. In addition, we observed an increased RD in APP/PS1 mice as compared with WT mice at 8 months of age. Compared with these DT metrics, however, greater significant differences were observed for the DK metrics. As such, compared with WT mice, APP/PS1 mice showed (1) higher MK values at 4, 6 and 8 months of age; (2) higher AK values at 4, 6 and 8 months of age; (3) higher RK values at 8 months of age; and (4) lower FA values at 4, 6 and 8 months of age. Additionally, we observed an age effect and an interaction between age and genotype for MK, AK and RK. For the retrosplenial cortex, the MK, AK and RK values showed an age effect, a genotype effect, and an interaction between age and genotype. However, the increased AK and decreased FA values of APP/PS1 mice compared with WT mice were the only straightforward differences which persisted over time and became larger as pathology progressed. Albeit to a lesser degree, this trend of increased MK and AK values in APP/PS1 mice compared with WT mice was also observed in the frontal association cortex. The AK values of the cingulate cortex did show an age effect, as well as an interaction between age and genotype, and were significantly increased at 4, 6 and 8 months of age in APP/PS1 mice compared with WT mice. In contrast, in the hippocampus, we did not observe any significant changes in DK metrics, but we did observe significant changes in the DT metrics. MD, axial diffusivity and RD were significantly increased in APP/PS1 mice as compared with WT mice at 6 and 8 months of age, and an interaction between genotype and age was observed for all three DT metrics. Next, from 4 months of age onwards, the piriform cortex showed significantly reduced axial diffusivity but increased AK in APP/PS1 mice compared with WT mice. In addition, an age effect and an interaction between age and genotype were noted for AK. For both the piriform cortex and the visual cortex, a significantly decreased FA value in the APP/PS1 mice was noted as compared with WT mice, respectively from month 6 and month 4 onwards. Additionally, for the FA values of the piriform cortex, we observed an age effect, a genotype effect and an interaction between age and genotype. In the septal nuclei, we observed decreased axial diffusivity in APP/PS1 mice as compared with WT mice, which was significant from month 4 onwards. More importantly, however, an age effect and an interaction between age and phenotype were observed in addition to a progressive decrease of the FA values in APP/PS1 mice as compared with WT mice, which were significant at all ages. Lastly, we observed a decrease in FA values in APP/PS1 mice compared with WT mice in all three parts of the corpus callosum: the genu, the body and the splenium. While this was significant at 2, 4 and 8 months of age in the genu, as well as at 2, 6 and 8 months of age in the body, the splenium showed a trend only towards decreased FA values in APP/PS1 mice.

Because the ROI-based analysis showed differences between APP/PS1 mice and WT mice in multiple ROIs, we investigated if we could successfully predict the genotype of the mice on the basis of statistical linear discriminant modelling of the present data. For this purpose, we limited our analysis to the motor cortex because this region showed the most statistically significant differences in the ROI-based analysis (see above), and the DK metrics of the motor cortex correlated well with the histologically determined pathology (see below). In Fig. 4a, we show the MCE (percentage of animals assigned to the wrong genotype) of the LDA at 2, 4, 6 and 8 months of age when using only the DT metrics (top row), the DK metrics (middle row) or the DT and DK metrics combined (bottom row). At 2 months of age, the MCE for the DT metrics (0.43), the DK metrics (0.52) or the DT and DK metrics combined (0.48) indicates the inability of all three metric combinations to separate APP/PS1 mice from WT mice. However, the MCE decreases with age for all three metric combinations, and as such, at 8 months of age, the MCE for the DT metrics is 0.16, and for the DK metrics, it is 0.20. However, when we combined DT and DK metrics, the MCE was 0.09. When comparing the MCE of these three metric combinations at 8 months of age, we could show that these are significantly different from each other with a p value < 0.0001 for any of the three possible comparisons (DT vs DK, DT vs DT + DK, and DK vs DT+ DK). Taken together, this clearly indicates the added value of acquiring DKI data and using all metrics estimated.

Because the combination of both DT and DK metrics gave the lowest MCE, we performed LASSO analysis to investigate which of these metrics were the most important for a correct classification of the genotype. In Fig. 4b, we show a heat map of all 100 iterations of the LASSO analysis, which were done on the data of the 8-month-old mice. On one hand, the MK and FA are the most important metrics for a correct classification of the genotype because they were involved in 100% of the iterations. On the other hand, AK, RK and axial diffusivity were involved in 60%, 59% and 70% of the iterations, respectively. In contrast, MD and RD were involved in just 20% of the iterations. These results clearly indicate that, despite all seven metrics contributing to a different degree, a correct genotype classification depended foremost on the DK metrics and FA.

Following the ROI-based analysis, we wondered if the changes observed in DKI metrics correlated with the Aβ-driven pathological changes in the tissue microstructure. Therefore, we sliced the full left hemisphere and collected brain tissue slides at a 150-μm intervals, allowing us to create 3D histological stacks of the performed staining. In Fig. 5, we show a representative close-up of the motor cortex for both WT and APP/PS1 mice at each of the four ages. While these images are only qualitative, it can already be appreciated that the Aβ-induced pathology progresses as APP/PS1 mice become older, as shown by an increased intensity of 4G8, GFAP and IBA1 staining (Fig. 5a, c and d, respectively) and a decreased intensity of the MBP staining (Fig. 5b).

Next, we statistically modelled the ability of all DT and DK metrics to predict the corresponding histological parameters in the motor cortex. In Fig. 6a, we show the correlation graphs of the MK, AK, RK and FA metrics with the %O.D. 4G8, GFAP, IBA1 and MBP. The graphs of the correlations of MD, axial diffusivity and RD with these histological parameters are provided in Additional file 3, and the r and p values of all these correlations are shown in Fig. 6b. MK, AK and RK are positively correlated with the %O.D. 4G8, GFAP and IBA1 in a highly significant manner. MK and RK also correlate with the %O.D. MBP, albeit to a lesser degree. FA shows a significant negative correlation with the %O.D 4G8, GFAP, IBA1 and MBP. In contrast, the MD, axial diffusivity and RD metrics do not correlate with the %O.D. 4G8, GFAP, IBA1 and MBP (with the exception of axial diffusivity, which correlates negatively with the %O.D. MBP). This indicates that DK metrics correlated with the underlying Aβ-induced pathology, whereas DTmetrics did not correlate with Aβ-induced pathology. In Fig. 6b, we also show the r and p values of the correlations between MK, AK, RK, FA, MD, axial diffusivity and RD and the GFAP and IBA1 anisotropy values. The latter is a readout for the ramification of GFAP-positive astrocytes and IBA1-positive microglia. MK, AK and RK were significantly negatively correlated with GFAP anisotropy, whereas FA and axial diffusivity were significantly positively correlated with GFAP anisotropy. In contrast, MD and RD did not correlate with GFAP and IBA1 anisotropy. We did not observe any correlations between DT or DK metrics and IBA1 anisotropy.

To compare the ROI-based analyses with a voxel-based approach (which is often used in human studies), we performed statistical parametric mapping, and the resulting images of the mice at 8 months of age are shown in Fig. 7a. We did not observe differences at 2 and 4 months of age, and we observed only very limited differences at 6 months of age (data not shown). Red voxels indicate an increased value in the APP/PS1 mice as compared with the WT mice, whilst blue voxels indicate a decreased value in the APP/PS1 mice as compared with the WT mice. At 8 months of age, mice showed reduced FA in the septal nuclei (close-up shown in Fig. 7b). We also observed reduced FA and an increase in RD in the corpus callosum.

Next, we determined how well DKI metrics predict the underlying pathology at a voxel-based level. As described earlier, we created 3D histological stacks of the different histological stains, which were then co-registered to the corresponding 3D T2-weighted MRI image and, via this intermediate step, thus also to the DKI images. In Fig. 7c, we show the motor cortex average multivariate cross-validated correlation values (of all 8-month-old WT and APP/PS1 mice combined) between the percentage of area stained of the four different histological stains and the DKI metrics (combining all DT, all DK, or the DT and DK metrics together). The cross-validated correlation values at a voxel-based level were low. The best cross-validated correlations were found between DT and DK metrics and the %O.D. MBP. Interestingly, combining DT and DK metrics provided improved prediction of the %O.D. MBP than either DT or DK metrics separately. This high cross-validated correlation value between the DT and DK metrics and the %O.D. MBP contrasted with the low cross-validated correlation values between the DT and DK metrics and the percentage of 4G8 staining, the %O.D. GFAP, and the %O.D. IBA1. This indicates inability to predict amyloid-induced pathology from the DT and DK metrics at a voxel-based level.