2.1.

Mice Groups

For vascular imaging, heterozygous APP/PS1 and their control littermates (C57BL/6J, here referred to as WT) mice were divided into three age groups (2-, 4.5-, and 8-months-old). Each group consisted of $N=5$ animals for a total of $N=30$. APP/PS1 mice start developing AD around 4.5 months with full pathology presenting at 8 months. Two other groups of 21 WT and 25 AD mice at 6 months were also used to gather perfusion MRI estimates in the same species.

2.2.

Brain Perfusion and Fixation

Animals were briefly anesthetized with 2% isoflurane and received intracardiac perfusion perfusion through the apex of the heart with 30-mL phosphate buffered saline (PBS, $0.3\mathrm{mL}/\min$) followed by 30 mL of 4% paraformaldehyde (PFA) in PBS also at a rate of $0.3\mathrm{mL}/\min$ and then perfused with an FITC-gelatin perfusate (20 mL at $0.6\mathrm{mL}/\min$, 2% porcine skin gelatin type A (G1890, Sigma), 1% fluorescein-labeled-albumin (A9771, Sigma) with the body tilted to 30-deg head down in supine position. At the last step, the mouse head was submerged in ice water with the heart clamped to ensure gel solidification. After 15 min of cooling, the brain was carefully retrieved and kept in PFA for later imaging.

2.3.

Imaging System

The imaging system used a femtosecond dual beam laser (Insight DS+, Spectra Physics) with tunable range from 680 to 1300 nm. A motorized $XY$-stage (Zaber T-Series LSR150B) allowed moving the sample for mosaic acquisitions, to cover the entire sample. The objective used was XLFluor $4\times /340$, NA 0.28, WD 29.5 mm (Olympus, Japan). Sample $Z$-positioning was controlled by a motorized labjack (MLJ050, Thorlabs). The scanning system consisted of galvo mirrors (GVS002, Thorlabs) coupled by two converging lenses to place both mirrors at the scanning distance of the scan objective to reduce beam “walking” on the imaging-objective back aperture. A $4\times$ beam expander was installed behind the galvo mirrors to fill the imaging-objective aperture with the telescope. A variable attenuator consisting of a polarizing cube and half-wave plate controlled the incoming power to the microscope. To ensure a constant signal throughout the tissue, laser power was controlled during the acquisition to correct for attenuation. A custom-made vibratome was installed to cut sample blocks. Fluorescence signal detection was done following a dichroic long pass (Chroma, 695dcxxr) and a bandpass filter (ET525/50M-2P, Chroma Technology) placed in front of the photomultiplier tube (Hamamatsu PMT). After the acquisition of a complete slice, a section of 200 μm was removed to expose a new surface to the microscope. Figure 1 shows an overview of the apparatus.

Fig. 1

Overview of the serial two-photon microscope. Legend: $B_1$: $-25\mathrm{mm}$, LC1054; ${\mathrm{B}}_2$: 100.0 mm, LA1207-B; $f_1$: 25.4 mm AC254-030-B Lens (Thorlabs); $f_2$: 400 mm focal AC508-400-B-ML Lens (Thorlabs); ${\mathrm{Obj}}_1$: LSM04-BB, working distance 42.3 mm; and ${\mathrm{Obj}}_2$: XL Fluor $4\times$ objective; M: mirrors.

Fig. 2

ESF profile (top left) showing 10% and 90% points on the original image (bottom left). The distance separating these points is $2.1\mu \mathrm{m}$. Side view of the PSF shows the elongated form given by the $4\times$ objective (middle). Top view of an imaged bead (top right). The profile of the bead long axis yields an FWHM of $25\mu \mathrm{m}$.

Imaging of the sample was implemented by combining images from different physical positions set by the $XY$ stages. The raw images were taken with a 20% overlap in the $XY$ plane and a 25% overlap in the depth $Z$-axis to ensure continuity in the stitching reconstruction. The microscope then proceeded automatically through the tiles, followed by vibratome sectioning, repeating the procedure until the whole brain was imaged. An algorithm was implemented to detect the cut sample surface using the maximal fluorescence signal received from a complete A-scan (moving the objective along $Z$). Measuring the surface at multiple locations, a plane was fit to the data and was used for image stack acquisition. As the vibratome did not cut a perfectly straight plane parallel to the imaging focal plane, this scan allowed modeling the imaging plane tilt along the entire sample to ensure that all the sample was imaged uniformly.

2.4.

Characterization of the Imaging System

To measure the axial resolution of the microscope, $4\text{-}\mu \mathrm{m}$ diameter fluorescent beads (F36909 FocalCheckTM fluorescence microscope test slide #1, ThermoFisher) were scanned. An imaging sequence with $z$-steps of $10\mu \mathrm{m}$ was performed and analyzed. To compute the lateral resolution of the system, a target (R1L1S1N, Negative 1951 USAF Target, ThorLabs) was used. The edge spread function (ESF) from an edge of the target lines was used for lateral resolution estimation with a digital step of $0.1\times 0.1{\mu \mathrm{m}}^2$ to ensure an image limited by optical resolution. See Fig. 2 for more details.

2.5.

Image Reconstruction

Reconstruction consisted in transformations on raw tile data, to form a full 3-D image. The raw data were labeled on a 3-D grid to perform a tile registration optimization using a phase-correlation pairwise registration method inspired by Preibisch et al.³⁵ and developed by Lefebvre et al.³¹ Stitching of the $XY$ planes was done using a Laplacian diffusion method with a blending width of 0.2 in units of overlap size fraction.³¹ After reconstruction, a maximum intensity profile (MIP) was applied on $250\text{-}\mu \mathrm{m}$-thick slabs. In some instances, we discarded unclear slices caused by issues described later in this report.

2.6.

Coregistration Method and Vascular Template Construction

As a first step, the imaged brains were digitally downsized to $25\times 25\times 25{\mu \mathrm{m}}^3$ resolution. Advanced normalization tools (ANTs) were used for coregistration to a template from the Allen Brain Institute using rigid followed by nonlinear registration.^36,37 As a second step, as a custom template typically outperforms using a generic template,³⁸ these coregistered data were used as input to an algorithm inspired from Refs. 39 and 40 to create a vascular template from our own data. By including the mirror images of the 30 brains, a grand total of 60 brains were coregistered to form a vascular template, which was also coregistered to the Allen template. The ROIs defined from the Allen Brain Institute mouse template were then deformed onto each brain through the transformations defined by the deformation fields, and used for evaluations below.

2.7.

Vessel Extraction and Volume Computation

Large vessels ($15\mu \mathrm{m}$ in diameter or larger) were segmented using a variable thresholding based on local image properties by a local Otsu threshold on image patches 1501 pixels or $3002\text{-}\mu \mathrm{m}$ large, followed by a small median filter ($2\times 2\text{pixels}$ or $4\times 4\mu \mathrm{m}$) to reduce segmentation noise. For smaller vessels, a Sato filter using a diameter of $2\mu \mathrm{m}$ was applied.⁴¹ As the latter method can only distinguish eigenvalues in the data, it can underline the contour of larger vessels causing a “webbing” artifact (see Fig. 5). Combining both filters resulted in a multiscaled image with contrast for both small and large vessels in the imaged organ. After the segmentation process, a manual verification was performed to correct major deformations that could have been due to artifacts such as leakage of the vascular system causing diffusion zones in the sample. After this last step, and keeping only microvasculature ($15\mu \mathrm{m}$ in diameter or less), a distance transform was applied on the mask followed by skeletonization. To compute vascular density, a model of cylindrical vessels using the pixel value as radii and the resolution as length was computed based on the distance map evaluated on the skeleton.

2.8.

Whole Brain Local Vascular Density Comparison

To investigate whether VD changes had a spatial structure, we estimated local VD in small-windowed regions of $200\mu \mathrm{m}$ and built reduced-resolution VD images ($200\times 200\times 200{\mu \mathrm{m}}^3$). These images were then mapped to the vascular template to ensure a correspondence between each voxel. Voxel-wise analysis was performed with a two-way analysis of variance (ANOVA) with age and the genetic strain as factors, using the SPM12 toolbox.⁴² The ANOVA results showed no interaction and a statistically weak main effect of the genetic strain over the whole cortex, consistent with the analysis of Cifuentes et al.²⁹ Therefore, the ROIs in Cifuentes et al., namely the hippocampal formation, the medial prefrontal cortex (mpfc), and cerebral cortex without the mpfc, were chosen to pursue the analysis, to which we added also the olfactory bulb based on the voxel-wise ANOVA.

2.9.

MRI Data for Comparison

To assess whether VD changes had MRI correlates, we compared our vascular density TPM results with perfusion acquisitions. Twenty-one WT and 25 AD mice all aged 6 months (midtime point compared to the 4.5 and 8 months groups used for TPM) were imaged in an MRI scanner. Anatomical and perfusion MRI scans were performed on a 30-cm 7T horizontal MR scanner (Agilent, Palo Alto, California) with mice in prone position, with a 12-cm inner diameter gradient coil insert, gradient strength $600\mathrm{mT}/\mathrm{m}$, and rise time $130\mu \mathrm{s}$. A two-channel receive only surface coil positioned over the mouse brain was used in combination with a quadrature transmit/receive birdcage coil with an internal diameter of 69 mm (RAPID Biomedical, Germany). Anesthesia was maintained with 1.5% to 2.5% isoflurane in 30% oxygen in air and body temperature was maintained at 37.0°C using a warm air fan (SA Instruments, Stony Brook, New York). Respiration ($\text{target}=100$, allowed range before adjusting $\text{isoflurane}=80$ to 120 BPM) and heart rate were monitored, the latter with a pulse oximeter.

An anatomical image was acquired with a 3-D true free induction with steady-state precession (TFISP) sequence⁴³ at $100\text{-}\mu \mathrm{m}$ isotropic resolution ($\mathrm{TR}=5.0\mathrm{ms}/\mathrm{TE}=2.5\mathrm{ms}$, 16 frequency shifts, 22-min scan time), used for coregistration. Then, a 3-D amplitude-modulated continuous arterial spin labeling scan (amCASL) was run,⁴⁴ with $\mathrm{TR}=3.0\mathrm{s}$, 1.0-s labeling duration, $60\times 54\times 48$ matrix, $300\text{-}\times 333\text{-}\times 333\text{-}\mu \mathrm{m}$ resolution, and 23 mins scan time. The tagging plane was kept in a fixed position nominally perpendicular to the carotids.

2.10.

MRI Data Analysis

Perfusion was calculated voxel by voxel using formula 1 in Ref. 44, with some parameters assumed constant: brain/blood $\text{partition coefficient}=0.9\mathrm{mL}/\mathrm{g}$, mouse arterial blood transit $\text{time}=0.08\mathrm{s}$, $\text{tagging efficacy}=0.67$, $T_{1\mathrm{b}}=2.3\mathrm{s}$, $T_1=1.53\mathrm{s}$, $T_{1\mathrm{sat}}=0.57\mathrm{s}$, and $M_a^z(\mathrm{w})=0.48\mathrm{s}$. The anatomical scans were first manually coregistered rigidly using ITK-SNAP⁴⁵ to a previously generated mouse template, then coregistered nonlinearly with ANTs.⁴⁶ This provided the ANTs transformations to coregister the perfusion images. Coregistration results were inspected in detail, and registration parameters were stored as potential confounding variables. SPM⁴² voxel-by-voxel analyses were run on the perfusion images. The lack of any statistically strong effect suggested pursuing the analysis using the above four ROIs as for the ex vivo scanner study. The MarsBar⁴⁷ toolbox in SPM was also used for the ROI statistics and to extract the average perfusion data.

3.1.

Optical System Characterization

As a first step, the two-photon microscopy system was evaluated using resolution targets described in Sec. 2.4. From fluorescent bead imaging and the ESF computation, the resolution of the optical system was measured to be $2.1\mu \mathrm{m}$ in the lateral direction, and had a full width at half maximum (FWHM) of $25\mu \mathrm{m}$ in $z$. This experimental resolution ensured that $50\text{-}\mu \mathrm{m}$ $z$-motor steps during the acquisition procedure led to independent data from neighboring voxels.

3.2.

Raw Data Coregistration and Image Segmentation

A vascular template was built from the data for each brain downsampled at an isotropic resolution ($25\mu \mathrm{m}$). In this process, each brain was coregistered to this new template. The template, and hence each brain, were also coregistered to the Allen template. Figure 3 shows a typical coregistration for one brain.

Fig. 3

(a) Representative $250\text{-}\mu \mathrm{m}$ MIP from one mouse, reconstructed from the two-photon microscope data. (b) Brain template from the Allen’s Brain Institute. (c) Overlay image representing the template (B in green) coregistered on the data (A in red) using the ANTs library.

Figure 4 shows the created template (B) aligned to the Allen template (A). Note that some sharp bright features are present in the vascular template. They were inspected to correspond to large blood vessels seen on many individual brains. We also observed that some regions of the olfactory bulb template have lower intensity. This corresponded to a few brains for which the bulb was incomplete. This may have arisen because all brains were cut along coronal sections with the olfactory bulb being the last surface to be exposed. Its mechanical stability in oxidized agarose was sometimes compromised, leading to data loss.

Fig. 4

Illustration of the Allen template with $25\text{-}\mu \mathrm{m}$ resolution (a) compared with the custom vascular template with the same resolution (b).

The ROIs derived from the Allen template atlas were mapped to the individual brains, using the transformation fields from template construction, and upsampled in the process to the original spatial resolution of the acquisition. These masks were then used for VD estimations derived from the segmentation process shown below.

The segmentation procedure described in Sec. 2.7 was applied on each brain using the full resolution. Figure 5 shows the results of this segmentation at intermediate steps, including the webbing effect in large vessels following Sato filtering and its correction through a mask combination step. The process yielded local estimates of vessel radius, which were used to compute VD with or without thresholds for vessel size.

Fig. 5

Visualization of the segmentation algorithm (a) MIP from one brain. (b) Zoom on the inferior left hippocampal complex. (c) Local Otsu threshold applied on B. (d) Sato filter applied followed by an Otsu threshold. One can see the “webbing” vessels pointed by the red arrow of larger caused by the Sato filter. The local Otsu threshold combined with a median filter captures the larger blood vessels, whereas the Sato filter recovers smaller vessels with little salt-and-pepper noise. (e) A distance transformation is performed on the combination of the two masks and (f) combined with the vessel center points derived from skeletonization by multiplication. (g) As a result, we obtain an image in which pixels are the vessels radii.

3.3.

Voxel-Wise Statistical Parametric Mapping Analysis

At $p<0.05$ uncorrected for false positives, results of the 2 ANOVA for VD showed a main effect of genetic strain covering the cortical area, the upper hippocampal complex, and some of the olfactory bulb. As discussed above, unfortunately, the olfactory bulb was truncated due to missing data for this structure and only the portion having complete overlap across all brains was kept for analysis. Based on this result, we selected the ROIs shown in Fig. 6 below.

Fig. 6

(a) Sagittal view of the Allen template with Student’s T-test positive for the contrast $\mathrm{APP}/\mathrm{PS}1>\mathrm{WT}$ at $p<0.05$, uncorrected, (b) axial view, and (c) coronal view.

3.4.

Regions of Interest Analysis

Guided by previous observations using sparse sampling histology and the voxel-wise analysis above, we evaluated VD in four predefined ROIs: the hippocampal complex, the medial prefrontal cortex, the (truncated) olfactory bulb, and the cortex without the mpfc (Fig. 7).

Fig. 7

(a) Sagittal view of the Allen template with the different ROIs (Hippocampal complex in blue, cortex in red, truncated olfactory bulb in cyan and medial prefrontal cortex in green). (b) Axial view of A.

Following the ROI mapping process and restricting to the microvasculature ($<15\mu \mathrm{m}$) only, VD was computed and is shown in Fig. 8 for each ROI.

Fig. 8

Vascular density for each group and ROI. (a) Hippocampal ROI, (b) medial prefrontal cortex ROI. (c) Cortex without the mpfc. (d) Truncated olfactory bulb. One-tailed student T-tests were performed with $*p<0.05$ considered significant after Tukey–Kramer correction for the contrasts and Bonferroni correction for the four ROIs. Error bars represent one standard error on mean.

Two-way ANOVAs with age and genetic strain as factors showed a statistical significance for the mpfc and the olfactory bulb for the age factor, after Bonferroni correction for four regions, and removing two mice that were outliers. Posthoc tests showed that the VD decrease in 8-months-old mice compared with 2 months was significant for the above two ROIs. There, we also observed a tendency for VD decrease for both the olfactory bulb and the medial prefrontal cortex between 2 and 4.5 and for all ROIs between 4.5- and 8-months-old for both APP/PS1 and WT mice. Furthermore, there was a tendency for higher VD in the APP/PS1 group when compared with the WT. The MRI perfusion data analyzed for comparison showed no statistical significance for the four ROIs. However, a nearly significant effect is present in the whole olfactory bulb ($p=0.017$, $\mathrm{APP}/\mathrm{PS}1>\mathrm{WT}$), correlating with a higher VD in the same region but that disappeared following Bonferroni.