Methods to label, image, and analyze the complex structural architectures of microvascular networks

Abstract Microvascular networks play key roles in oxygen transport and nutrient delivery to meet the varied and dynamic metabolic needs of different tissues throughout the body, and their spatial architectures of interconnected blood vessel segments are highly complex. Moreover, functional adaptations of the microcirculation enabled by structural adaptations in microvascular network architecture are required for development, wound healing, and often invoked in disease conditions, including the top eight causes of death in the Unites States. Effective characterization of microvascular network architectures is not only limited by the available techniques to visualize microvessels but also reliant on the available quantitative metrics that accurately delineate between spatial patterns in altered networks. In this review, we survey models used for studying the microvasculature, methods to label and image microvessels, and the metrics and software packages used to quantify microvascular networks. These programs have provided researchers with invaluable tools, yet we estimate that they have collectively attained low adoption rates, possibly due to limitations with basic validation, segmentation performance, and nonstandard sets of quantification metrics. To address these existing constraints, we discuss opportunities to improve effectiveness, rigor, and reproducibility of microvascular network quantification to better serve the current and future needs of microvascular research.


| INTRODUC TI ON
The microvasculature plays a plethora of key roles in maintaining tissue homeostasis, including modulating oxygen transport, 1 nutrient delivery, inflammation response, 2 and wound healing. 3 Structural changes to the microvascular architecture have been shown to profoundly regulate these fundamental biologic processes. 4 Therefore, characterization of the complex changes in spatial structure of the microvascular architecture gives a better understanding of the roles microvessels play in pathogenesis, maintenance, prevention, and amelioration of diseases. Indeed, the importance of the microvasculature has long been appreciated in diseases such as small vessel disease, 5 coronary microvascular disease, 6 and the abundance of complications associated with diabetes. 7 However, recent research has indicated that the microvasculature also plays key roles in the top eight causes of death in the United States 8 ( Figure 1) and many others, including (1) heart disease: impaired infarct wound healing, 9 reduced oxygenation, 10 pulmonary hypertension in pre-capillary and post-capillary vessels 11 ; (2) cancer: pathological angiogenesis, 12 enriched microvessel permeability, 13 significant route for metastasis 14 ; (3) lower respiratory disease: capillary dropout, 15 reduced muscle oxygenation, 16 airway rigidity from vasodilation 17 ; (4) stroke: impaired microvascular flow patterns 18 and reduced oxygenation, 19 pericyte constriction of capillaries, 20 dropout of functioning capillaries 21 ; (5) unintentional injuries: angiogenesis, 22 clot formation, 23 immune cell recruitment 24 ; (6) Alzheimer's: attenuated vasodilation response, 25 amyloid angiopathy, 26 and tissue hypoxia 25 ; (7) diabetes mellitus: capillary permeability, pericyte dropout, capillary dropout 27 ; and (8) pneumonia and influenzas: capillary permability, 28 immune cell recruitment, 28 and impaired lung oxygen transport. 29 Additionally, the microvasculature is recognized as one of the most promising routes of drug delivery 30 by enabling direct targeting of microvascular endothelial cells with intravascularly injected drugs to exert profound therapeutic effects in disease conditions. 31 The overall import of the microvasculature in biomedical research is quickly approaching that of the nearly ubiquitous roles that the immune system plays in basic organismal processes 32,33 and disease development, 34 and future research focused on microvascular structure, function, and adaptations promises profound opportunities for curing human disease.
In this review, we highlight new key developments and survey contemporary and classical models of the microvasculature, along with techniques to label and image microvessels at high resolution where the complete microvascular structure is captured. Therefore, microvascular research that fails to resolve the smallest-sized vascular structures is omitted or given less emphasis, such as fundus imaging of the retina 35 and other clinical imaging methods. Although a subset of the modalities covered can yield 3D images, we focus on analysis of 2D projections of 3D vessel networks since it can be universally applied to all microvascular imaging modalities, 2D representations of 3D networks retain much of their information, 36 and 2D methods for quantification of vessel architecture can be extended to three dimensions, 37 although we do comment on the potential pitfalls of using 2D metrics to characterize 3D microvascular structures. Furthermore, many of the 3D modalities for microvascular imaging have reduced axial resolution compared to lateral, and practical considerations of acquisition time usually lead to further reduced axial sampling. 38 The currently available programs to analyze and quantify microvascular structures are also covered, along with constructive proposals for improvement in this area. While each topic covered could be a detailed examination on its own, this review is meant to offer a basic orientation of the technological options available for microvascular research and a perspective on analytical techniques to increase scientific rigor as science faces an ongoing crisis in reproducibility. 39

| MARKER S AND MODEL S TO S TUDY MI CROVA SCUL AR NE T WORK ARCHITEC TURE
The study of the complex architecture of microvasculature requires proper labeling and visualization of microvessels, using either a marker for particular cell types, unique basement membrane constituents, and/or labeling perfused microvessels via the intravascular injection of a dye or fluorescently tagged antibody to visualize blood flow through microvessel lumens. [40][41][42] All of these labeling methods provide a means of contrasting microvascular architectures with the surrounding tissue when paired with a suitable imaging modality. The particular choice of vascular marker and imaging method should be carefully evaluated based on the biological questions being examined and determined based on requirements for resolution, signal to noise, tissue penetration depth, imaging location in terms of in vivo/ex vivo, and labeled cell specificity. 38 The relative importance of the various labeling and imaging considerations for microvascular visualization depends on the nature of the research and biological questions asked. For example, with investigations focusing on angiogenesis and subsets of vessel types, cell labeling of specific subpopulations is essential, while for studies characterizing blood flow, accurate vessel diameter and connectivity between vessel segments have greater significance. Moreover, effective pairing of these technologies with a particular imaging approach requires an understanding of the fundamental strengths and weaknesses of the options available.

| Markers of microvasculature
A critically important aspect of studying blood vessels is carefully tailoring biological interpretations and conclusions to appropriately correspond to the specific cell types or structures visualized (Table 1). An example that illustrates incongruity between data and interpretation is when vessels are labeled via F I G U R E 1 The significance of the microvasculature in top causes of death and disease in United States. Top eight classes of fatal disease or injury with the fraction of annual deaths in the United States. Included with each malady are three highlighted fundamental roles the microvasculature plays with initiation, maintenance, or treatment (see main text for references) perfusion of fluorescent dye 43 and general conclusions are made about vascular remodeling, disregarding changes in structure of nonperfused vessels, vessel neosprouts, and regressing vessels. 44 Even with specifically worded conclusions, focusing on findings that only quantify portions of the microvascular architecture represents an incomplete analysis and may omit significant remodeling events. Another key example is the use of superfused IB4 lectin, a marker commonly used to label blood vessels that also labels pericyte 45 and macrophages 46 (Figure 2A). Especially in development, many papers prematurely conclude changes occur in vascular architecture based on lectin staining 47,48 while omitting the issue that a mix of cell types are labeled, especially with the inability to differentiate structures between pericyte and endothelial cells. Studies that use Col-IV staining to quantify microvascular architecture 49 have similar shortcomings, labeling not just blood vessels 50 ( Figure 2B), but other cell types such as pericytes and fibroblasts. 51 Additionally, Col-IV also marks thin bridges between capillaries previously referred to as Col-IV sleeves, 52 string vessels, 53 and acellular capillaries 54 in various tissues such as retina, 54 brain, 55  is interpreted as a sign of collapsed or regressed vessels, yet this has never actually been established. There is a clear separation between the two structures in thickness ( Figure 2C) and with the cross-sectional pixel intensity profile between lumenized vessels and Col-IV tracks ( Figure 2D)  are ASMA-negative, 61 but if the samples were snap frozen with methanol fixation to prevent actin depolymerization, at least half of capillary pericytes are ASMA-positive. 61 We suspect that major portions of other canonically known ASMA-negative pericyte populations across tissues might actually express this marker, and there is a possibility that ASMA is, in fact, a pan marker for pericytes that requires a more sensitive measurement technique to confirm.
However, even if ASMA is expressed by all pericytes to some degree, it is not a unique marker for pericytes, because it is also expressed by vascular smooth muscle cells.
Finally, the expression of Tie2 by pericytes has been fiercely debated in the past decades, with extensive characterization of Tie2 expression in cultured pericytes, 62 but a lack of Tie2 expression noted in pericytes in vivo. 63  which measures RNA expression of the target gene directly.

| Animal models with endogenously labeled vasculature
An increasing number of transgenic murine models have been developed to visualize the microvasculature, including those that contain cell-type specific fluorescent reporters for endothelial cells, smooth muscle cell, and pericytes (Table 2). We emphasize there is an important limitation of these reporter models that is often ignored: these animal models often only include the proximal endogenous reporter region with the fluorescent reporter, meaning that gene expression behavior from distal enhancers is often lost. An example of this is with Tie2 expression, which has been found in other cell types such as HSCs. 66 and neutrophils, endothelial progenitor cells, macrophages, 67 pericytes, 62 and keratinocytes. 68 Yet the primary Tie2-GFP mouse model is only known for GFP expression exclusively in endothelial cells, 69 in this case serving as an advantage with a reporter line that appears to selectively label the vasculature and not track other cell types known to have endogenous expression.

| In vitro and ex vivo models to study the microvasculature
Over the last century, various models have been used to study the complexity of the microvasculature, including those that utilize TA B L E 2 Animal Models to Label Microvasculature cultures of various cell populations, tissue explants, and animal models (Table 3). Historically, researchers have had to consider significant trade-offs when choosing between different model systems. In vivo models typically have the best chance of recapitulating human disease since pathologies are heavily influenced by the complex interplay of a multitude of cell types. 70 However, this benefit comes at a cost: In vivo models are usually limited by throughput, exhibit high cost in time and resources, and imaging techniques restricted by limitations with sedation duration, subject restraint, and detector scan speed. Furthermore, typically any intervention short of a cell-specific knockout of an implicated gene will not establish cellular mechanism, which can take years to generate in an in vivo model. In vitro models typically exhibit much higher throughput and have a wider range of available analysis tools to characterize the system, 71 but at the cost of greater simplification and abstraction of tissue structure and disease conditions compared to in vivo, such as lacking various cell types or blood flow.
However, the trade-offs between in vivo and in vitro models are blurring now more than ever. Advances in new imaging techniques allow for in vivo imaging that provides the opportunity for higher throughput and fully temporal measurements in various tissues. The latest in vivo gene editing techniques, such as CRISPR/Cas9, 72

Gluteus Maximus
In vivo In vivo Ms 232 , Rt 233

Microfluidic EC Chip
In vitro

EC Microbeads in Fibrin
In vitro Var. 248  research, and patient-specific primary cell culture allow for more advanced in vitro models, although there is still difficulty with cell collection in these systems for subsequent analysis. 71 Indeed, the number of available model systems has been growing, and with the advent of new analysis techniques, the opportunities to collect data from microvascular network architecture have increased dramatically and reveal new prospects for efficient and reproducible data capture.

| STATE-OF THE-ART IMAGING MODALITIES FOR MICROVASCUL AR NETWORKS
There are a vast range of techniques available for imaging the microvasculature, with trade-offs between resolution, signal penetration, and acquiring multiplexed functional readouts, such as

| Metrics for quantifying microvascular networks
Previous research has developed various metrics for microvascular network analysis ( Figure 3A-G), including the fraction of image area composed of blood vessels (VAF), 77 81 and max extra-vascular diffusion distance to examine tissue oxygen perfusion. 82 Other metrics have been developed outside of this set but not standardized and adopted by consensus. Studies often normalize metrics in different ways, such as measuring branchpoints per image, normalizing to field of view, or normalizing to vessel length.
We posit that metrics should be designed to encourage valid comparisons across research studies and should be normalized to facilitate this process. Thus, using a simple metric of vessel length 83 is not as useful as vessel length density, a metric that can be directly compared over a range of spatial resolutions and imaging modalities. This tissue environment heterogeneity is further reflected by unique endothelial transcriptomes found in each organ 85 and distinct endothelial marker profiles at different parts of the vascular tree. 86 With much of biology, function and form are closely inter-

| The need for pairing perivascular and microvascular analysis
Analyzing changes to the perivascular space can yield just as important insights as the vasculature itself given the close cross-talk between endothelial cells and smooth muscle cell and pericytes. For example, changes in pericyte density are known to play a key role in the pathogenesis of diabetes, 27 and changes in pericyte locations relative to branchpoints 96 have been associated with changes to stability and sensitization of the microvascular network. We emphasize the need to analyze perivascular behavior as well as microvascular remodeling to truly understand the structure and function of micro-

| Software packages for quantifying microvascular networks
Alterations in microvessel network architecture have been used ubiquitously in studying vascular diseases, and there are a multitude of software packages available for quantifying changes in architecture (Table 5). Three, notably, have been used in a significant number of publications, namely AngioQuant, AngioTool, and RAVE.
AngioQuant has been developed to analyze endothelial networks in vitro, with a focus on quantifying various metrics of tubule formation using bright field images. 83 Recently, it has been adapted for use in evaluating higher resolution datasets of microvascular networks in vivo 105 and in histological samples. 106 Its validation is focused on quantification of in vitro experiments showing trends of changes with various metrics, but no statistical comparisons between study groups.
The datasets were not validated against manually analyzed images and there is no analysis included comparing accuracy and overall performance between AngioQuant and other available software packages.
AngioTool is presented as a quick, hands-off, and reproducible image analysis tool, deployed as an ImageJ plugin, for quantification of microvascular networks in microscopic images. 78 The validation of AngioTool included analyzing biological data from murine hindbrain and retina using various metrics, including visualized vessel segmentation, vessel centerline, and branchpoints presented for qualitative inspection. For quantitative biological validation, endothelial cell explants were cultured and analyzed with two drug treatments that were known to alter vascular structure as a positive control. 78 Additionally, output metrics were validated with a subset of manually counted images in an unblinded fashion with two investigators.
RAVE is an image analysis tool that can be used on a wide array of images 81 to accelerate the unbiased, quantitative analysis of F I G U R E 6 Novel image analysis metrics by analyzing the microvasculature with graph theory. A, Confocal microscopy image of the murine retinal deep vasculature, with CD105 marking ECs (white; scale bar = 50 μm). B, Image analyzed with basic segmentation, skeletonization, and branchpoints classification, with vessel centerline/ skeleton (white), branchpoints (red), and EP (turquoise). C, Conversion of vasculature into a graph, with branchpoints indicated as "nodes" (blue) and vessel segments as "edges" (gray). D, Visual summary of types of metrics to quantify graphs, with removed edges representing change to network after a blood vessel has regressed or experiences obstructed flow

| Improving vessel quantification, analysis, and interpretation
We estimate that these software packages are largely underutilized, based on the high number of published manuscripts that refer to the quantification of microvessel architecture. Indeed, a search on PubMed on relevant terminology (terms used included: microvasculature density, capillary dropout, pericyte dropout; see all terms in Appendix S1) reveals over 120 000 publications to date. While this query includes publications that merely mention the terms searched for, the nearly three order-of-magnitude difference between citations of these software programs compared to this large collection of publications suggests that there is an unmet need for vessel architecture analysis beyond the available options, with researchers often resorting to manual ad hoc analysis of microvessel networks, leading to decreased repeatability, comparability, and scientific rigor.
We propose the following design criteria for an effective software package: • • Effect of parameter adjustment: Some software packages allow for adjustment of key image processing parameters to enhance results, but the effect (and bias) of allowing the user to freely alter image processing outcomes needs to be rigorously examined and reported.
• Blinded image analysis: Software packages should include built-in support for image filename anonymization to blind the user from an image's study group assignment to minimize bias as images are analyzed.
• Semi-automated curation: Image quality and marker expression can change between study groups, which could bias automated Without tackling these issues, new software programs are merely presented to research scientists "as is" without allowing them to make informed decisions on how to produce high-quality unbiased results.
We argue that until these issues are dealt with, the use of these packages has the risk of leading to a significant error rate in microvascular research: where the software reveals a positive error with a quantifiable change between groups where none existed, or even worse, where research is not pursued based on a negative error where no change is observed between study groups where one exists.

| Applications of machine learning, graph theory, and modeling in quantifying microvascular architecture
Metric effectiveness not only needs to be evaluated on an individual basis, but the effectiveness of a given collection of metrics in combination needs to be evaluated. This can start with examining covariance matrices of metrics from in silico and biological datasets to evaluate how much unique information they bring relative to one another. Effectiveness of metric sets could be evaluated using principal components analysis, partial least squares, or more advanced methods of feature selection techniques from the field of feature engineering, 108 117 and we believe there are a wealth of relevant metrics that could be applied from quantifying graph networks ( Figure 6).
Examples include metrics measuring centrality of each node, 118 scoring the relative importance of edges in connecting nodes, connective redundancy, 119 and information diffusion. 120 Machine learning methods such as convolution neural networks and other techniques 121 have also been applied to graph networks. 122 Previous applications of network tomography outside of biomedical research, 123 such as internet tomography, 124 may be a pertinent source of applicable methods for characterizing complex microvascular networks.

| Relevant techniques from related network architectures
Analysis of full feature microvascular architectures can also benefit from adapting techniques used to analyze similar network structures. A prime example is fundus imaging of human retinal vascular networks: Although such images fail to resolve capillary portions of the microvascular network, this application has been extensively explored due to its established clinical significance in evaluating eye disease and a collection of systemic diseases, 125 with a plethora of methods developed to segment, quantify, and validate vessel architecture. 126 Indeed, a machine learning-based classification pipeline for fundus images has recently been approved by the FDA for diagnosis of diabetic retinopathy. 127 Methods used in the processing and quantification of neuronal network image datasets 128 may yield useful techniques for analyzing microvascular networks. 129 In vivo clinical imaging techniques, such as micro-CT 130 of lung vascular networks, may also yield insight into extending 2D imaging and quantification into the third dimension. 131 Although this review focuses on 2D image quantification techniques, many of the imaging modalities mentioned acquire data in 3D directly, or through a series of 2D slices. Over the long term, the field would most benefit from acquiring and analyzing 3D datasets to eliminate any confounding phenomena that arises from analyzing a 2D-projected representation of complex 3D structures. Such 2D abstractions can lead to altered metrics, such as false branch points where vessels appear to overlap in 2D but exist at distinct elevations in 3D, introducing error to other metrics such as segment length.
Furthermore, the 3D orientation of vessel segments relative to one another is especially important when characterizing local tissue oxygenation. 89 An in-depth examination of how 2D structural metrics can characterize a projected 3D structure like a microvascular network would also be necessary to understand the trade-offs and reveal what information is missed with this simplification.

| Statistics to analyze microvascular interactions: beyond generic
The metrics covered in this review require basic 2-sample or multisample statistical tests to determine whether there is a difference in structure and morphology of vascular networks between study groups. Yet there are also new statistics being developed based on modeling null distributions (output if the null hypothesis is true and there is no difference between groups) that could be extended to quantifying the microvascular architecture. A prime example of this is a technique to measure cellular recruitment with a given cell type and the vascular network, 132 that maintains validity in conditions where generic statistics fail. While cell recruitment has been analyzed, 97 previous metrics of cell-to-cell colocalization events fails to properly measure changes in cell recruitment if there are changes to vascular density or cell density across study groups. This confounding phenomenon will lead to false positives when testing between study groups 132 : instances where the test indicates there is a significant change in cell recruitment, when in reality there is none. Null modeling of random cell placement is used to avoid the deficiencies of generic statistics, and we believe this modeling approach could be applied to evaluating perivascular cell recruitment to blood vessel architecture using an in silico model and provide researchers with a more robust statistical hypothesis test for analyzing microvessel architecture.

PER S PEC TIVE S
The microvasculature is implicated in pathogenesis and maintenance of the deadliest maladies of the modern world.
Understanding microvasculature's function, adaptation, and contribution to disease is enabled by the application of metrics that quantify changes to microvascular network architecture in in vivo, in vitro, and in silico model systems. We highlight opportunities to further the field by improving scientific rigor and reproducibility through the development and validation of software that reliably, comprehensively, and in an automated manner, characterizes the complexities of microvascular architecture using pre-existing and novel metrics.