1.

Introduction

Adequate blood flow is essential to supporting healthy brain function.¹ Global and regional reductions^2–6 in cerebral blood flow as well as alterations in hemodynamic^3,6 responses are associated with or precede cognitive decline across multiple neurodegenerative diseases, including Alzheimer’s and Parkinson’s diseases. At the smallest scale, alterations in microvascular structure⁷ and flow can also drive pathology,^8–10 including regions affected by stroke induced hypoxia. This includes transient disruptions in flow, or “stalls,” in normally continuous cerebral capillary flow.^11–13 Recent work has demonstrated the potential role of stalls in pathology, including Alzheimer’s disease¹⁴ and stroke.^15–18 In addition to potentially inducing local hypoxia, these events can contribute to total vascular resistance,¹⁴ as well as flow heterogeneity,¹⁹ further inducing hypoxia.²⁰ Treatments to reduce this stalling have been associated with favorable cognitive outcomes in a model of Alzheimer’s¹⁴ and reduced infarct size in stroke.^15,18

Investigations into the effects of stalling and its potential as a therapeutic target have been hindered due to technological limitations. Traditional two-photon imaging $z$-stacks are frequently used to measure stalling rates^12,14–17 but only observe individual capillaries for a small fraction of the acquisition time. Further, even when a stall is observed, the onset and cessation of events is typically not seen. The alternative is to monitor a single $XY$ plane over a long duration, but due to two-photon confinement, only a handful of capillaries are typically visible in a single plane. Optical coherence tomography (OCT) angiography provides continuous volumetric monitoring of capillaries^13,18 but the OCT systems used to study stalling have lacked the temporal resolution to catch shorter stalling events [though faster OCT angiography (OCTA) rates have been demonstrated²¹ but not yet applied to measuring stalls]. In addition, OCT also lacks the ability to perform multicolor fluorescence imaging like two-photon microscopy, which has aided in investigating the source of the stalling events.^14,15,17,18

In addition to these limitations, analyzing and extracting stalling information for these enormous datasets is highly labor intensive. For both two-photon microscopy and OCT, images must be manually analyzed to identify the stalling events. In the case of two-photon microscopy, analysis must be crowd sourced to citizen scientists who volunteer to sift through hundreds of thousands of capillary images to identify stalled capillaries so that the resulting statistics can be reported in a reasonable amount of time.^14,22 Further, the sparse nature of stalling in these data makes the construction of a large training dataset for automation difficult due to class imbalance. As more complex investigations are performed and therapeutic modulation of stalling continues to be explored, the amount of data that needs analysis will grow significantly. Therefore, there is a need for high throughput imaging and analysis of stalling parameters to avoid a data processing bottleneck.

To address these challenges, we use a custom Bessel beam two-photon microscope²³ to identify stalling events by continuously monitoring a $713\times 713\times 120\mu \mathrm{m}$ volume acquired at 0.57 Hz, roughly four times the volumetric imaging rate of conventional OCT angiograms. We then developed a correlation-based detection approach to semi-automate the analysis for stall identification, which reduced the time required to analyze the data by 50% and led to a 2 to 3 times improvement in the number of stalls identified when compared with unassisted analysis. With the combined improvements in temporal resolution and improved detection rate, we found that a majority of stalls are very short in duration, making it likely they were missed in other studies. This enabled us to observe differences in stall dynamics between young and aged mice as well as track changes in stall dynamics before and after stroke in the same capillary network.

2.

Results

We used a home built, axicon-based, Bessel beam two-photon microscope to image cortical vasculature over an extended depth of field²⁴ [Fig. 1(a)]. Within this volume, roughly 200 capillaries are monitored for both short- and long-duration stalling events [Fig. 1(b) and Video 1], where red blood cell (RBC) shadows are stationary for at least two or more consecutive frames. Identification of stalled capillaries from 10-min time courses, and the times they were stalled are used to calculate key stall parameters¹³ [Fig. 1(c)]. This includes the percent of all capillaries that undergo stalling at any time (incidence), the average percent of capillaries actively stalling at a given time (point prevalence), and the average total time a stalled capillary spends stalling during the measurement duration (cumulative stall duration). Identification of stalling capillaries and annotating the time and duration of each stall event for the acquired time series data is a particularly laborious task since there are hundreds of capillaries to observe at higher volume rates than previous studies. To make stalls more apparent, the data is broken down into sub-ROIs containing $\sim 20$ to 30 capillaries each, which is then scrolled through frame by frame to identify instances where all RBC shadows within a capillary are stationary between two or more frames. This process is repeated for all sub-ROIs until the whole field of view has been processed. A full $713\times 713\times 120\mu \mathrm{m}$ dataset of 350 frames took close to 12 h per dataset and the results can be variable depending on the focus and fatigue level of the observer. To improve both throughput and detection accuracy, we introduced a metric that can predict the occurrence of a stall in each capillary.

Fig. 1

(a) Maximum intensity projection (MIP) of a $713\times 713\times 120\mu \mathrm{m}$ volume scanned at 0.57 Hz for 10 min to observe instances of capillary stalling. (b) Representative examples of types at stalling events captured during imaging. (Top) RBC shadows are stationary for two sequential frames during a short stalling event before normal flow resumes. (Bottom) A shadow is continuously present in the same position along the length of the capillary during a long stalling event. (c) MIP with stalled capillaries marked in red. (d) Stallogram shows the times each capillary was stalled (marked in black). (e) Stall statistics calculated based on the stallogram and total number of capillaries (Video 1, mp4, 7.54 MB [URL: https://doi.org/10.1117/1.NPh.10.3.035009.s1]).

During normal flow, at the scanning rate used, new shadows (mostly RBCs) will be present in every frame, varying in number and position. During a stall, the same shadows will remain in the same position along the length of the capillary until flow resumes in that capillary. Therefore, the intensity along the length of the vessel is expected to be well correlated during stalling events. To facilitate extraction of the intensity along the length of the capillary, we semi-automatically extract the centerline trace of capillaries of interest²⁵ [Fig. 2(a)]. This generates images of intensity along the capillary length versus time (LT images). Plotting the frame-to-frame intensity correlation versus time shows a clear increase during stalling events, before returning to lower levels [Fig. 2(b)]. To validate this metric, we threshold the correlation and use it as a predictor of stalling, which was then confirmed by a human observer.

Fig. 2

(a) Capillaries that undergo stalling with the centerline drawn automatically after single click selection. (b) LT plots of the capillaries from (a) with the correlation versus time plotted alongside. Red boxes indicate times where the capillary was stalled and the same RBCs are present for multiple frames, which corresponds to a rise in frame-to-frame correlation.

2.2.

Comparison to OCT Stalling

After validation of our stall analysis, we then compared our resulting stall parameters using the correlation aided detection to stalling statistics previously reported with OCT (Table 2).¹³ We found stalling incidence was twice that reported by OCT, likely due to the increased temporal resolution of the Bessel beam two-photon microscope. When looking at the distribution of the duration of the stalling events [Fig. 3(d)], most of the individual stalling events lasted $<8\mathrm{s}$ and therefore would easily be missed OCT-based angiography, which takes $\sim 8\mathrm{s}$ to generate an image. This also contributes to the lower cumulative stall duration we measured with the Bessel beam microscope, as OCT was only sensitive to longer stalls and thus does not identify many capillaries that have only short duration stalling events [as suggested by Fig. 3(a)]. To confirm this, we removed short duration stalling events in our Bessel beam results and then recalculated the statistics. When we only consider stalls occurring for $\sim 8\mathrm{s}$ or longer, our measured incidence is very similar to what was found previously with OCT [Fig. 3(a), Table 2]. The cumulative stall duration [Fig. 3(c)] did remain lower, still likely due to the increased temporal resolution and improved ability to identify the end of longer stalling events with the Bessel beam approach. The variability in stall point prevalence found in OCT [Fig. 3(b)] was too high to make a precise comparison to that found with the Bessel beam results, but the results are none-the-less consistent.

Table 2

Comparison of stalling statistics measured by OCT (reported in Ref. 13) and by Bessel beam two-photon microscopy (BB-TPM) in two different age groups (4 to 8 months and 14 months). The long stalls only column calculates stall statistics for stalls lasting 5 Bessel frames or more (i.e., ∼8  s) to facilitate comparison with the previously published OCT results.

	OCT13 (∼7 months)	BB-TPM (4 to 8 months)	Long stalls only	BB-TPM (14 months)	Long stalls only
ROIs	12	10	10	14	14
Animals	6	4	4	4	4
Minimum stall duration (s)	$\sim 8$	1.75	$\sim 8$	1.75	$\sim 8$
Incidence (%)	6.8 ± 2.9	13.02± 6.2	5.24 ± 2.74	27.5 ± 12	8.4 ± 4.16
Point prevalence (%)	0.31 ± 0.7	0.58 ± 0.3	0.38 ± 0.23	1.18 ± 0.6	0.49 ± 0.37
Cumulative stall duration	10.6 ± 3.2	4.49 ± 1.97	6.91 ± 3.39	4.29 ± 1.49	6.4 ± 5.17

Fig. 3

(a)–(c) Stalling statistics calculated after removing stalls shorter than a given length. Stall length of two includes all stalls detected. Purple point indicates the same statistics reported previously by OCT for a chronic window¹³ used to compare (a) stall incidence, (b) point prevalence, and (c) cumulative stall duration. (d) Cumulative distribution function of all individual stalling events of young and aged groups. Dashed line represents the time to generate a single OCT angiogram. (e) Stall incidence as a function of measurement time. Dashed black lines shows fit of the exponential model $I_{\text{stall}}=A(1-\exp (-\frac{B}{t}))+C)$. The best fit resulted in $A=8.4$, $B=6.4$, and $C=1.15$ for the aged group and $A=5.4$, $B=6.5$, and $C=0.7$ for the young group.

We then applied our correlation aided approach to analyze data taken from older ($n=4$ mice, 14-month-old) mice and compared it to the results from the younger group analyzed in the validation study. The older cohort had increased rates of stalling compared to the younger group, predominantly of shorter duration [Fig. 3(d)]. When only looking at longer duration stalls (more than 4 frames, i.e., close to $\sim 8\mathrm{s}$), stall incidence was closer to the young cohort and previous OCT results, though still elevated. In both cases, the incidence of these long stalling capillaries over measurement time can be fit to a simple exponential model $I_{\text{stall}}=A(1-\exp (-\frac{B}{t}))+C)$, indicating a subset of capillaries consistently stall rather than being randomly distributed over the entire capillary network.

2.3.

Photothrombotic Stroke

To demonstrate the ability of our system to track changes in stalling statistics, we used a targeted photothrombotic model of stroke²⁶ and measured the differences in capillary stalling at baseline compared with 1 week post-stroke in 14-month-old mice. Regions-of-interest (ROIs) were determined to be in the stroke core, peri-infarct, and contralesional hemisphere as identified using spatial frequency domain imaging (SFDI) to track changes in tissue scattering [Fig. 4(b)].²⁷ In total, there were three ROIs in the stroke core, three in the peri-infarct region, and three in the contralesional hemisphere across two animals. Both ROIs in the stroke core and all but one ROI in the peri-infarct region saw increased stalling incidence. The only ROI that did not show an increased incidence at week 1 also had the highest incidence at baseline. We also found an increased presence of longer stall durations in all ROIs, including the ROIs in the contralesional hemisphere [Figs. 4(e) and 4(f)]. The stalling point prevalence was also increased [Fig. 4(d)], primarily driven by stalling duration, as it increased even in regions where overall incidence was unchanged. This shift is also apparent when looking at the distribution of the durations of the stalling events across all capillaries [Fig. 4(f)]. At baseline, the older mice had higher incidence, which was primarily driven by shorter stalling events. One week after stroke, longer duration events account for a larger percent of all stalls, even in the contralesional control ROI [Fig. 4(g)].

Fig. 4

(a) Widefield reflectance image showing cortical vasculature at baseline. Red arrow indicates the artery that was targeted for photothrombotic stroke. Boxes show the ROIs that were imaged at baseline and 1 week post-stroke. Colors indicate the relation of the ROI to the stroke. Red indicates stroke core, black peri-infarct (peri), and blue contralesional hemisphere (control). (b) SFDI reduced scattering coefficient (${\mu }_s^{\prime }$) and relative images show the increase in scattering as a result of the stroke, used to determine the stroke core. (c)–(e) Changes in stall statistics in 7 ROIs across 2 mice from baseline to 1 week after stroke. Color indicates the relation to the stroke core. (f) Cumulative percentage of stalling events across all stalling capillaries as a function of duration across different groups. Line indicates the percentage of stalls with a duration less than or equal to the duration. (g) Cumulative percentage of stalling events across stalling capillaries based on region at 1-week post-stroke (Video 2, mp4, 11.1 MB [URL: https://doi.org/10.1117/1.NPh.10.3.035009.s2]).

2.4.

Arterial Dilations Transiently Reduce Stalls

In addition to volumetric imaging offered by the extended depth of field, Bessel beam allows for simple estimation of vessel diameter using the vessel fluorescent intensity²³ [Fig. 5(b)]. We hypothesized that periods of increased flow may reduce stalls by providing sufficient pressure across the capillary to clear the stall. We therefore analyzed the average stalling point prevalence around large, isolated dilation events (see Sec. 4). We found that the average point prevalence was significantly reduced by the time of peak dilation [Fig. 5(c)] and shortly after. This decline started prior to the peak diameter because flow starts to increase as soon as vessel diameter starts to increase.

Fig. 5

(a) Full stalling field of view with an artery on the left-hand side. Scale bar is $100\mu \mathrm{m}$. (b) Diameter and fluorescence intensity ($\mathrm{\Delta }F/F$) trace of a section of the artery shows close correlation between calcium fluorescence and diameter with the Bessel beam (correlation coefficient: 0.88). Black circles indicate time points used for analysis in (c). (c) Changes in stall point prevalence (black) relative to large dilation events (red), averaged across 9 ROIs. Time 0 corresponds to the time of peak diameter during dilation. Error bars indicate the standard errors estimated across $n=65$ dilations from 9 ROIs.

3.

Discussion

To overcome the limited volume rates of conventional OCTA and two-photon angiography, we used a custom-built Bessel beam two-photon microscope to monitor a $713\times 713\times 120\mu \mathrm{m}$ volume at 0.57 Hz for instances of stalled capillary flow. This was 3 to 4 times faster than OCT angiography used in the previous studies to monitor a similar volume.^13,18,28,29 The increased speed led to a bottleneck in analysis, which was previously done fully manually. We semi-automatically extracted intensity along the centerline of all capillaries of interest and used it to flag time points of interest in all capillaries. The high false positive rate meant that our correlation metric was not sufficiently robust to directly detect stalls, but it was enough to reduce the time burden of analysis by about half (from 12 to $\sim 6\mathrm{h}$). It also identified the vast majority of stalling events, including many that blind observation missed. The latter is especially important as the results of the fully manual analysis indicated that observers missed as much as 70% of all the stalling events (Table 1), and that this manual analysis was adapted from the approach used to analyze OCT angiograms.^13,18 This is understandable given the complexity and size of even a single dataset and given that it is easier to confirm that the automatically identified stalling events are truly a stalling event as opposed to the fully manual approach, which requires blind inspection of all capillaries at all time points. A similar metric could also be adapted to assist with the analysis of OCT angiograms.

While we were able to greatly reduce the analysis burden while improving accuracy, stalling analysis was still time-consuming. Deep-learning-based tools have shown promise in distinguishing these events from normal flowing capillaries,³⁰ which could further reduce or eliminate the need for human observers. Due to the increased rate of stall detection with our approach compared with traditional two-photon microscopy, the relatively small number of datasets used in our studies still contain processed results from over 6000 capillaries that were monitored for 10 min each. Therefore, a sufficiently large training set of stalled and flowing capillaries could be created with fewer datasets compared to previous acquisition methods.³⁰ Our analysis would also likely benefit from automated extraction of capillary traces from the Bessel beam data.^30–32

We also leveraged the ability of Bessel beam to monitor pial artery dynamics simultaneously with capillary stalling [Fig. 5(a)]. As has been shown by others, vessel diameter correlates highly with vessel intensity²³ [Fig. 5(b)], as a larger volume of intravascular dye is excited when the vessel diameter is increased. This provides an easy to calculate metric for vessel diameter that is insensitive to axial motion, which we used to observe drops in stalling point prevalence around large dilation events [Fig. 5(c)], as stalls were presumably cleared due to the transient increase in flow and perfusion pressure. This is consistent with recent work showing that environmental enrichment reduced the number of capillaries plugged by injected microspheres.³³ But here we show that there is on average fewer stalled capillaries around larger arterial dilation events in vivo, demonstrating the potential for dynamic reductions in stalling from spontaneous flow changes, not just under direct functional activation.¹³ This effect was shown with large and isolated dilation events, but more sophisticated analysis can be done to better understand the relationship between stalling and neurovascular coupling, or even stalling and resting state oscillations.³⁴

To demonstrate the ability of our approach to detect changes in stall characteristics, we compared the statistics between young and old mice to previously reported statistics.¹³ Similar to the previous work, we found that a subset of capillaries repeatedly stalled [Fig. 1(d) and Table 2], rather than being randomly distributed across the capillary network.¹³ The similarity in long stalling parameters [Table 2 and Fig. 3(e)] to OCTA gives confidence that we are observing the same events with the addition of shorter events OCTA could not capture previously. At baseline, the older mice had more than twice the number of capillaries that stalled at any given time (point prevalence) and in total (incidence) but had a larger share of stalling events that were shorter in duration [Fig. 4(f)] resulting in a lower cumulative stall duration. This is further confirmed when removing shorter stalling events from analysis, resulting in a stall incidence more similar to the younger group. The stall incidence of longer stalling events was consistent with that reported with OCT,¹³ where slower frame rates were used and thus would not be sensitive to shorter events. In the older group, two mice also underwent a photothrombotic model of stroke and were reimaged 1-week post-stroke. Stroke resulted in a significant rise in stalling point prevalence and cumulative stall duration. This indicates that there is an increase in the duration of stalling events, even in the contralesional hemisphere. When recomputing the stall statistics using only longer events, an increase in stall incidence is also observed (Fig. S1 in the Supplementary Material). Circulation of neutrophils can still be elevated at this 1 week stage in recovery,³⁵ which could drive increases in stalling across the cerebral vascular network, including in the contralesional hemisphere. However, a larger scale study would be needed to confirm these phenomena. Future work can also address the source of longer versus shorter duration stalls using fluorescent labeling methods. For example, neutrophils adhering to the vessel wall could be the main driver of longer stalls as well as explain why only a subset of capillaries repeatedly stalls. This would be consistent with the previous work, where administering Anti-Ly6G to interfere with the adhesion of neutrophils reduced stalling.^14,18 The question of whether neutrophils primarily drive longer stalls could be addressed with multicolor imaging using Rhodamine 6G as a second fluorophore^12,14 (Video 2).

Our ability to detect shorter stalls also raises the question of the shortest stall duration that should be considered during analysis. Previous work was only sensitive to longer stalls due to the nature of the measurement approaches and showed that the reduction in long stalls was critical and increased total cerebral blood flow well in agreement with a computational model.¹⁴ In particular, it was found that a decrease in stalls from 1.8% to 0.72% of all capillaries resulted in a 13% increase in total cerebral blood flow.¹⁴ Similar modeling and analysis have not yet been done to account for brief stalls in flow, but modeling of capillary transit times has shown that overall flow heterogeneity in the capillary network reduced oxygen delivery efficiency, even for the same total cerebral blood flow.^19,36 Increased stalling, even for brief periods, would lead to increased intracapillary flow heterogeneity and therefore be potentially detrimental despite their short duration. There is also predicted upstream and downstream effects of flow stoppages such that even brief stalls could theoretically increase spatial flow heterogeneity,³⁷ meaning flow heterogeneity is created across the local network. The increase in short stalls we found in our aged mice would be consistent with the increased spatial capillary speed heterogeneity found in older mice, where pockets of hypoxic tissue were also found more frequently.³⁸ On the other hand, while brief stalls are likely driving increases in flow heterogeneity that could lead to pockets of hypoxic tissue, the diffusion of oxygen from the capillary tissue acts as a low-pass filter of capillary oxygen dynamics³⁹ and smooths out brief drops in oxygen, such that brief stalls might not induce critical levels of hypoxia in the adjacent tissue. Therefore, more work needs to be done to determine the impact of varying stall durations on oxygen delivery.

Since the imaging volume rate is equivalent to the imaging frame rate with Bessel beam two-photon microscopy, it offers the potential for imaging capillaries at very high-volume rates.²³ The frame rate used in our study was relatively slow compared to what is achievable with even standard linear galvanometers but was used to allow for large field of view images with a high signal–to-noise ratio, ensuring a large number of capillaries can be monitored. It was also important to keep frame periods sufficiently long such that the presence of the same RBCs between frames [Fig. 1(b)] was indicative of stalled flow, as even small shifts in RBC position would still mean a flow speed on the order of $\mu \mathrm{m}/\mathrm{s}$ instead of mm/s as is typically observed.^11,40,41 Using higher frames rates (such as 100 to 1000 Hz^23,42,43) where RBC flow can be tracked would require a speed cutoff for stall classification. Further work needs to be done to understand what duration of stall is required to significantly affect oxygen delivery, which would then inform the optimal temporal resolution or speed threshold for quantifying their dynamics. Regardless, future investigations into capillary stalling will greatly benefit from high speed, high throughput measurements of capillary stalling statistics. Bessel beam two-photon microscopy offers an ideal platform with an improved volume rate over conventional approaches and multicolor imaging of fluorescently labeled cells and plasma that can easily monitor hundreds of capillaries to study these events.

4.

Methods