Near-lifespan longitudinal tracking of brain microvascular morphology, topology, and flow in male mice

In age-related neurodegenerative diseases, pathology often develops slowly across the lifespan. As one example, in diseases such as Alzheimer’s, vascular decline is believed to onset decades ahead of symptomology. However, challenges inherent in current microscopic methods make longitudinal tracking of such vascular decline difficult. Here, we describe a suite of methods for measuring brain vascular dynamics and anatomy in mice for over seven months in the same field of view. This approach is enabled by advances in optical coherence tomography (OCT) and image processing algorithms including deep learning. These integrated methods enabled us to simultaneously monitor distinct vascular properties spanning morphology, topology, and function of the microvasculature across all scales: large pial vessels, penetrating cortical vessels, and capillaries. We have demonstrated this technical capability in wild-type and 3xTg male mice. The capability will allow comprehensive and longitudinal study of a broad range of progressive vascular diseases, and normal aging, in key model systems.

after Benjamini-Hochberg multi-hypotheses correction). The correlation was calculated between two time-courses of every possible pair, with age lags of 0, 4, 8, and 12 weeks. Pairs with positive and negative correlations are shown in red and blue respectively. More opaque lines indicate higher maximum correlation, and thinner lines means that the maximum correlation appeared at longer age lags. A correlation with an age lag means that one alteration was in (a). Scale bar, 0.1 mm. c) 17 slices were analyzed (12 slices from 3 AD mice and 5 slices from 2 WT mice). All GFAP images underwent an identical image segmentation process, which used adaptive thresholding and filtered foreground objects by the area (equivalent to circles of 10-30 µm in diameter). DAPI images underwent circle detection processing. The differences were normally distributed, and we used paired, two-sided t-tests to analyze them. We excluded one slice as an outlier in the area ratio analysis. The effect of AD on this conclusion of insignificant  13 slices were analyzed (9 slices from 2 AD and 4 slices from 1 WT mice). All IBA-1 images underwent the identical image segmentation as that for GFAP images. DAPI images also underwent the identical circle detection processing.
The differences were normally distributed, and we used paired, two-sided t-tests to analyze them.
The effect of AD on this conclusion of insignificant difference was not statistically significant (p=0.093, number ratio; 0.054, area ratio; LME analysis).
Supplementary Figure 10. Effects of isoflurane anesthesia on individual vessels. a) Effects of isoflurane on pial vessel diameter. 61 vessels (young) and 59 vessels (old) were measured and analyzed by bootstrapped LME, as traditional LME analysis produced non-normally distributed residuals. No multiple-comparison correction was needed. The bootstrapped LME, especially when bootstrapping the measures within each group, does not determine an exact p-value. b) Effects of isoflurane on arteriolar and venular diameter and flow. 11 arterioles and 15 venules (young) and 16 arterioles and 17 venules (old) were measured and analyzed by LME. Residuals were normally distributed. No multiple-comparison correction was needed. Each box chart displays the median (thick line), the lower and upper quartiles (box), and the minimum and maximum values that are not outliers (whiskers, computed using the 1.5x interquartile range). Gould et al. 6 Haft-Javaherian et al. 7 Ji et al. 1 Blinder et al. 4 Kirst et al. 2 Branching order

Supplementary Text 1. Novelty statement
A variety of OCT imaging and image-processing techniques have been developed, adopted, and integrated for the presented framework. It ranges from reproduction of published methods (e.g., blood flow measurement from Doppler OCT data) and novel application of existing concepts (e.g., betweenness of a capillary network as adopted from graph theory) to development and validation of novel methods (e.g., the deep-learning method for capillary RBC flux measurement; Figs. 4a and 4b). But more importantly, the major novelty of the presented work lies in the integration of these existing and newly developed techniques into a single framework which has enabled to produce new types of information that otherwise would be highly challenging to obtain if not impossible (e.g., Figs. 5c and 5d).

RBC flux measurement.
Our CNN was also designed to provide a measure of uncertainty in each prediction, allowing us to filter out high-uncertainty predictions in further analysis. In contrast, the traditional peakcounting method does not provide the uncertainty of its estimation; thus, we used a single color on the left figure of Fig. 4a. This information about uncertainty was one of the many advantages of using the CNN-based method against the traditional peak-counting method. Even without this uncertainty filtering, the CNN outperformed the traditional peak-counting method when tested on data unseen by the CNN during training ( Fig. 4a; the slope is closer to 1, and R 2 is higher).
When filtering out predictions with the lowest 20% confidence, the error became both smaller (narrower distribution) and less biased (the mean closer to 0 RBC/s) even for higher flux values ( Fig. 4b).
Regarding two-photon microscopy (TPM) data as the ground truth, TPM has been used as a standard method of measuring RBC flux and speed in capillary vessels of the rodent brain cortex [8][9][10][11][12][13] . It is based on fluorescence (staining plasma in most cases) and thereby yields a high signal-to-noise ratio. It is also capable of detecting RBC passage from vessels located deep in the cortical tissue, generally down to hundreds of micrometers deep, which cannot be achieved by using an older approach based on video microscopy [14][15][16] . Thus far, to the best of our knowledge, no techniques offer more accurate measurements of RBC flux and speed than TPM. A related challenge in a longitudinal imaging experiment with the cranial window is a limitation in employing behavioral tests to simultaneously perform on the same set of animals. In this study, we employed the NOL test alone for several reasons. First, the Morris water maze test, widely used in this context, could pose a risk of failure due to potential water damage to the cranial window headpost. Second, fear conditioning could be a potential addition to the NOL test, but repeatedly subjecting the same animal to conditioning and testing phases for several months could raise animal welfare concerns. Additionally, applying electric shocks to an animal with a metal head post attached could be a potential issue. Finally, the 3xTg model used in this study is one of the most widely used transgenic models of AD, and its cognitive decline has been well characterized in the literature. For example, Belfiore et al. found spatial learning and memory deficits at 6 months of age (26 WOA) 17 , which is highly consistent with our result from the NOL test (27 WOA, Fig. 5b).

Supplementary
Another related challenge may be large variations in measurements between animals and/or ages.
If one compared absolute values between groups age by age from the presented data, it might suffer from lower statistical power. Such age-by-age comparison does not exploit the analytic advantage of longitudinal data ( Supplementary Fig. 1). Therefore, careful selection of a statistical method and clear focus of analysis are required. The LME method used here is often preferred for longitudinal data analysis over repeated ANOVA, as LME is relatively robust against missing values and violations of sphericity which increase rates of Type I error 18,19 .
Another potential challenge in robustness is that the absolute values of diameters obtained from OCT angiogram (OCTA) may be inaccurate because the mechanism of OCTA contrast originates from the movement of RBCs through vessels. This may make the OCTA-measured diameter vary with measurement positions along the vessel direction. To minimize such variations, we measured the pial diameter from the mean of several cross-sections extracted along the vessel centerline. But this improves precision, not accuracy. Furthermore, isoflurane used in this study is known to influence vessel diameter. Despite these limitations in accuracy, high precision is often sufficient for many studies that focus on relative differences between experimental conditions rather than absolute values, like in the presented study. For a study where measurement of the absolute vessel diameter is critical, methods like fluorescent twophoton microscopy may be more suitable although its long-term longitudinal imaging capability has not yet been demonstrated to the best of our knowledge.
In addition to the properties measured in this study, a few related OCT measures have recently been proposed by others, and they can be readily added to the presented framework of methods.
For example, capillary stalling has been found to play a role in pathology development in AD by a two-photon microscopy study 20 and then demonstrated to be measurable by OCT 21 . Also, it may be possible and interesting to study whether and how shear-induced diffusion of RBCs as measured by OCT 22 alters with aging.
Another potential opportunity of the presented method is to investigate subcortical vasculature when being used with a longer-wavelength OCT. We used the OCT with the center wavelength of 1.3 µm as it covers all cortical depths and the scope of this study focuses on cortical microvasculature. OCT with longer wavelengths like 1.7 µm, however, can provide more homogeneous depth penetration and potentially reduce scattering, enabling further investigation of subcortical vasculature 23 . Combining the presented method with such a longer-wavelength OCT could enable researchers to explore how the vasculature beneath the cortex gradually alters in animal models of Alzheimer's disease, and how the alteration differs from that of cortical vasculature.
A direct methodological parallelism is another interesting potential of the presented methods.
The OCT upon which the presented methods are based is already being widely used in clinical ophthalmology. Under the assumption that microvascular degenerations appear both in the brain and retina, some preclinical findings enabled by the presented methods may be directly applicable to human studies. In particular, a recent study suggests that cerebral microvascular degeneration may be one of the etiologies, independent of Aβ, in human ApoE4 AD 24 . ApoE4 accounts for a larger AD population than the familial AD for which the 3xTg model used in the present study was designed. Therefore, if the presented methods are applied to the ApoE4 AD model and produce a similar type of findings demonstrated here (e.g., the betweenness alters the earliest among capillary structural properties), and if such properties are measurable from human retinal OCT images, it would provide unprecedented potential for non-invasive, inexpensive, and imaging-based biomarkers for preemptive prediction of AD. Furthermore, such an application of the same method to both human and mouse will allow us to directly test any preclinical findings, a kind of ground truth comparison to animal models typically unavailable with most methods. Supplementary Text 5. Differences in the age of significance between arteriolar/venular diameter/flow.
The differences in the age of significance observed in Fig. 5c do not necessarily mean that arteriolar flow, for example, did not change while its diameter altered in early ages. As defined in the main text, the age of significance rather indicates the age at which a vascular alteration

Supplementary Text 7. Interpretation of capillary length and flow results.
Our results showing a decrease in mean capillary length with aging in AD (Fig. 3d) are consistent with a previous study that found capillary length to be shorter in AD mice than in wild-type (WT) mice at 18-31 WOA 7 , although that study did not longitudinally track the length and used a different AD mouse model. Another study found that capillary branch numbers were lower in 3xTg mice than WT mice at 20 months of age, but no difference was observed at 7 and 14 months 37 . This finding is consistent with our result of shorter capillary lengths in 3xTg mice, but our measurement revealed the difference much earlier (Fig. 3d). There are several possible reasons for this discrepancy, including that (i) the statistical analysis in the previous study did not consider the clustered nature of the data, leading to lower sensitivity to intra-group differences; (ii) our method tracks the trend of the capillary property varying with age, whereas their method relies on a snapshot at a specific age; and/or (iii) our method analyzes 3D networks, while their method analyzes only 2D cross-sections of the network.
The pattern of blood flow through a capillary vessel network plays an important role in the energy supply regulation of the brain and its malfunction in diseases [38][39][40][41] . In our results, RBC flux gradually increased with aging in WT mice, being consistent with previous findings from WT rats 42 although the previous study measured RBC flux at only two ages in a non-longitudinal manner. AD mice also showed an increase in RBC flux but at a higher rate (6.0% versus 2.1% per month, Table 1). We also observed reduction in the COV of RBC flux in AD mice (i.e., more homogenized capillary flow pattern). Whereas higher RBC flux and lower flow heterogeneity in capillaries are generally associated with higher oxygen extraction 11,38 , the observed increases in RBC flux and decreases in its heterogeneity with aging might be due to a compensatory mechanism for capillary vessels to suppress negative effects of the observed, earlier reduction in arteriolar and venular flow, similar to an adaptation occurring in hypoxia 43 . We speculate that such compensatory mechanisms do not last indefinitely; thus, if we had tracked the capillary flow properties for an extended period, we might have seen a decrease in capillary RBC flux and increase in RBC flux COV at a later age. Further studies are required to test this hypothesis.
The combined results of arteriolar diameter/flow and capillary flow suggest another possible consequence of decreased arteriolar diameter and flow, namely an increase in the presence of hypoxic micro-pockets in the cortex. The observed decreases in both arteriolar diameter and flow would lead to arteriolar and near-arteriolar tissue oxygen pressure (pO2) 44 . This impaired arteriolar oxygen delivery would not immediately lead to an impairment in capillary-bed tissue pO2 when the compensatory mechanism by enhanced capillary flow works as hypothesized above from our observed capillary RBC flux increases. However, when the compensatory mechanism no longer works at later ages, the capillary-bed tissue pO2 would become lower and spatially more heterogeneous, as observed between 60 and 100 WOA in WT mice 44 . This impaired capillary oxygen supply increases the presence of hypoxic micro-pockets, which may explain the observation of tinier microinfarcts in aging brains, particularly those with mild cognitive decline or AD 45,46 . Therefore, the decrease in arteriolar diameter and flow observed in both WT and 3xTg mice with age may have important implications for the development and progression of neurodegenerative diseases.
The decrease in penetrating arteriole and venule diameters ( Table 1) further suggests that the blood pressure might have increased with age in the 3xTg mice we used. Unfortunately, we did not longitudinally measure blood pressure in our study. Hypertension has been linked to AD, and animal studies suggest that hypertension can lead to amyloid plaques, neuroinflammation, bloodbrain barrier dysfunction, and cognitive impairment 47 . However, human studies on the association between hypertension and AD have been sparse and inconsistent, and the effect of antihypertensive medications on AD seems weak 48,49 . Investigating interactions among plaques, tangles, cerebrovascular pathology, and dementia may be key to understanding hypertension's role in AD development 49 . Therefore, our future work includes performing a long-term trace of blood pressure using a noninvasive method like the tail-cuff technique in diverse AD models, which will enable us to compare the long-term blood pressure trace with the traces of cerebral microvascular properties obtained by the presented methods.

Supplementary Text 8. Potential impact of long-term craniotomy and imaging on animal physiology.
As described in the Methods section, we started with 20 animals, but one animal did not survive until the end of our seven-month longitudinal experiment. Three animals were euthanized in the middle of the study due to mechanical damage to their headposts, and three animals were Although the spontaneous death rate and existing literature suggest that long-term craniotomy is unlikely to have adverse effects on animal physiology, there has been no direct investigation into this possibility. To address this, we conducted post-mortem immunofluorescence imaging of the brains of animals that had undergone our longitudinal experiment. Using the published method of Heo et al. 54 , we fixed, sectioned, and stained the brains with GFAP and DAPI ( Supplementary   Fig. 9a) or IBA-1 and DAPI ( Supplementary Fig. 9d), which respectively visualize astrocyte and microglia immunoreactivity. These two stains are widely used in the literature to investigate the effect of craniotomy on cortical physiology [51][52][53][54] . We measured astrocyte immunoreactivity by calculating the ratio of GFAP to DAPI in terms of cell number or pixel area, and found no statistically significant differences between the ipsilateral and contralateral cortices of the window installation in both ratios ( Supplementary Fig. 9c). We performed similar analysis on the DAPI and IBA-1 stained images and found no statistically significant differences in microglia immunoreactivity ( Supplementary Fig. 9f).

Supplementary Text 9. Impact of anesthesia on OCT measurements of vascular morphology, topology, and flow.
The presented study conducted OCT imaging under isoflurane anesthesia, which is known to dilate cerebral vessels. While this vasodilation effect may vary between imaging sessions, it would work as random noise and not affect the main conclusion of the study. This is because the study focused on comparing relative changes in the rate of change with age (RCA) between AD and WT groups, rather than absolute values. However, if the vasodilation effect is systematically different between younger and older mice, it could affect the accuracy of the RCA measurement.
Even in this case, it would not change the main conclusion of the study, as the focus was on the statistical difference in RCA between the groups, rather than absolute RCA values.
Nonetheless, it would be valuable to identify which vascular properties are affected by isoflurane anesthesia in an age-related manner, and which are not. As no study has comprehensively investigated this for younger and older mice, an additional experiment was conducted in WT mice.
Methods: Seven days after installing a chronic cranial window on a mouse as described in the main text, we placed the mouse on an air-floating platform (Mobile Homecage, Neurotar) for a head-fixed, freely-walking awake imaging condition, while connecting the mouse nose to the isoflurane vaporizer. Under this awake condition, we acquired a similar set of OCT data as in the main study, including OCT angiogram, Doppler OCT, microangiogram, and RBC passage data.
Then, we turned on the isoflurane supply (1.5% isoflurane with oxygen flow of 1 L/min), and after 30 minutes, we repeated the OCT data acquisition under this anesthetized condition. We repeated this experiment in young (15 WOA) and old (55 WOA) mice (n=4 mice per age group).
This dataset underwent the same analysis as described in the main study to measure all the vascular properties listed in Table 1. For statistical analysis of properties tracked vessel by vessel ( Supplementary Fig. 9), we used the LME method involving animal-specific random effects, excluding any outliers when a residual was more than three scaled median absolute deviations from the median. When this LME method resulted in non-normally distributed residuals, we used a bootstrapped LME method with a bootstrapping number of 2,000. Returning to the focus of this additional experiment, whether the isoflurane effect is different between young and old mice, the effect size of the older age was statistically significant for pial vessel diameter only (Supplementary Table 3, Supplementary Fig. 9a). The isoflurane-induced increase in pial vessel diameter was smaller in old mice (p<0.05).
Discussion: The isoflurane-induced vasodilation effect on pial vessel diameter was greater in younger mice. Therefore, when tracking the pial vessel diameter with aging and measuring its slope (RCA) as we did in the main study, a negative RCA may be observed even if the animals do not exhibit physiological decreases in pial vessel diameter. As expected, our main study observed a small negative RCA in pial vessel diameter, with no statistically significant difference between AD and WT mice (-1.3%/month and -1.1%/month for AD and WT, respectively, p=0.78 between AD and WT, Table 1). This negative RCA is highly likely due to the age-dependent pial vessel dilation effect of isoflurane. The effect size of the older age was -15% on average (Supplementary Table 3 Although the statistical results presented in the main text support that our approach was sensitive enough to reveal subtle differences in those vascular changes with age between AD and WT mice, we conducted a separate experiment to directly quantify the sensitivity of our approach of detecting changes in the average diameter and blood flow of pial vessels and penetrating arterioles. Methods: We used isoflurane as a vasodilator and repeated OCT imaging while vessels went through dynamic changes via the dilation and recovery. In detail, seven days after we installed a chronic cranial window on a mouse as described in the main text, we put the mouse on an airfloating platform (Mobile Homecage, Neurotar) for a head-fixed, freely-walking awake imaging condition, while the mouse nose is connected to the isoflurane vaporizer. We acquired nine sets of OCT angiogram and Doppler OCT images, following the exact protocols described in the main text, at the following time points: This longitudinal dataset underwent the analysis process as described in the main text, in order to identify same vessels across time points and measure the pial vessel diameter, arteriolar diameter and flow, and venular diameter and flow, for each vessel and each time point. We repeated this data acquisition and analysis in four young WT mice. For statistical analysis of a change between two states, we used linear mixed-effect (LME) analysis with the nested random effects of individual vessels in animals. This nested LME was identical to the LME analysis used in the main text.
Results: As expected, the pial vessel diameter increased with isoflurane and returned to the baseline after recovery (Supplementary Fig. 12a). Our approach was sensitive enough to detect a detectable 7% change in the pial vessel diameter between the first awake state and the state 5 minutes after turning on the isoflurane supply. The average pial vessel diameter in the awake state was 19.7 µm, and the increased diameter in the second state was 1.4 µm (p=0.026, n=59 vessels from four animals; Supplementary Fig. 12b). Our approach was also robust enough to produce statistically insignificant results between the first and the final states as physiologically expected (p=0.932, Supplementary Fig. 12c) while still detecting the 1.4-µm change.
Similarly, the arteriolar diameter and flow increased with isoflurane and returned to the baseline after recovery, while the venular diameter and flow did not ( Supplementary Fig. 12d). We found subtle fluctuations during the recovery phase, the degree of which our approach was able to detect. The average arteriolar diameter in the first of the final two states (45 minutes after turning off isoflurane supply) was 27.5 µm, and the increased diameter in the final state was 1.6 µm, a 6% change (p=0.022, n=7 arterioles from four animals; Supplementary Fig. 12e, top). The average arteriolar flow changed by 0.098 µL/min from 0.409 µL/min, a 24% change (p=0.052, Supplementary Fig. 12e, bottom). Our approach was also robust in producing statistically insignificant results between the first and the final states as biologically expected (p=0.942 and 0.436 for diameter and flow, respectively; Supplementary Fig. 12f).
Discussion: The supplementary experiment presented here provides a quantitative measure of the sensitivity of our approach for detecting changes in vascular diameter and blood flow. We found the sensitivity of detecting changes in pial vessel diameter, arteriolar diameter, and arteriolar blood flow to be 1.4 µm (7% change), 1.6 µm (6% change), and 0.098 uL/min (24% change), respectively. It is important to note that these sensitivity values serve as a conservative estimate of the actual sensitivity of the measurements described in the main text. This is because the supplementary experiment used fewer animals (four, against 6-7 per group of the main study) and compared only two time points, whereas the main experiment employed seven time points to determine the slope (rate of change with age, RCA). It is likely that the RCA measurement described in the main text has higher statistical power and thereby higher sensitivity.

Supplementary Text 11. Comparison to Aβ pathology in AD.
Various cerebral microvascular degenerations (CMDs) observed in this study became apparent between 12 and 25 WOA (Fig. 5c), preceding extracellular Aβ deposits in 3xTg model mice. In this specific model, extracellular Aβ deposits first become apparent in the frontal cortex at 6 months (26 WOA) and then evident in other cortical regions and in the hippocampus by 12 months (52 WOA) 27 , and cortical Aβ plaques were first detected at 12 months of age (52 WOA) 17 . This temporal relationship between microvascular and Aβ pathologies in the 3xTg model agreed with findings from a recent study using the corrosion cast method 55 . It is difficult to directly compare results between this terminal study and our study, because the terminal study compared absolute values between AD and WT mice, age by age, while our longitudinal study compared the slope of relative changes (RCA) between AD and WT considering all ages of measurement. Nevertheless, we found some consistency. In the somatosensory cortex, the area we investigated, the terminal study found the total length of capillary vessels (5-10 µm in diameter) was higher at 3 months of age, similar at 6 months, and lower at 12 months in 3xTg compared to WT mice, which is consistent with our result of decreasing capillary length in 3xTg mice (Fig. 3d). The terminal study also found no difference in vessel segment and vessel junction numbers between 3xTg and WT mice across the ages of 3-24 months. This result is agreeable with our result of insignificant differences in capillary branching order, number density, and length density (Table 1). Finally, the terminal study observed more tortuous vessels in 3xTg mice, although it did not conduct quantitative analysis of tortuosity, and we also found an increasing capillary tortuosity with age in 3xTg mice (Table 1).