Mathematical modelling of blood perfusion and oxygenation in microvascular networks with applications in stroke research Paul Sweeney September 2014 University College London Department of Mechanical Engineering Supervised by Dr Rebecca Shipley & Dr Simon Walker-Samuel Submitted in part fulfilment of the requirements for the degree of Master of Research in Mechanical Engineering of University College London
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Mathematical modelling of blood perfusion and
oxygenation in microvascular networks with
applications in stroke research
Paul Sweeney
September 2014
University College London
Department of Mechanical Engineering
Supervised by Dr Rebecca Shipley
& Dr Simon Walker-Samuel
Submitted in part fulfilment of the requirements for the degree of
Master of Research in Mechanical Engineering of University College London
Abstract
Stroke is the sudden onset of paralysis caused by either a bleed or blockage within
a brain’s vasculature, which prevents transportation of oxygen to brain tissue (brain
hypoxia). Stroke victims can recover with no or few disabilities if they are treated
promptly. Understanding the processes of fluid and mass transport behind this illness
could allow for the development of more effective methods of treatment and improve
our fundamental understanding of the disease state. Diagnosis of stroke itself is clinical,
with assistance from medical imaging. However, there is no commonly used blood test
for diagnosis. Control of blood pressure has conclusively shown to prevent ischaemic
stroke [39], although once a ischaemic stroke is diagnosed, it is not clear whether lower
blood pressure is helpful towards treatment. Currently two options are available for
therapy, breaking down the blockage or by removing it mechanically.
The aim of this report is to use and develop a mathematical model which simulates
blood perfusion and oxygenation in microvascular networks. Simulations are run using
the vascular data of a mouse brain gathered by CABI† using medical imaging. A com-
parison of network fluid pressure and tissue oxygen distributions in healthy networks
versus one containing an ischaemic stroke at the microvascular level will be presented.
Imaging is a ever developing research area, but current techniques can already extract
blood vessels down to a resolution of 5 µm. With access to these incredibly detailed
data sets on vascular structure, modelling can provide a tool to extract functional in-
formation.
Chapter 1 gives a brief introduction to brain anatomy, with focusing turning towards
current models of hemodynamics based on in vivo and in vitro blood rheology descrip-
tions [31, 33, 34, 38]. Chapter 2 focuses on the development of an algorithm to retrieve
the largest subnetwork from an incomplete network dataset (a method required due to
an incomplete vascular casting), with later discussion turning to the implementation of a
geometry-based algorithm developed by Smith [49]. Chapter 3 presents the framework
The subject of stroke is prevalent in today’s society where 1 in 6 people around the world
will suffer from a stroke in their lifetime [1]. In the UK alone, there are approximately
152,000 strokes every year, with 1 in 5 being fatal [1]. Stroke is a sudden and devastat-
ing illness but by combining mathematical modelling with medical imaging, we aim to
better identify and treat damaged sections of the brain and improve our fundamental
understanding of the mechanical and transport processes at play.†
The aim of this thesis is to use and combine existing mathematical models which
simulate blood perfusion and oxygenation in microvascular networks. Simulations will
be run on a healthy network and a ischaemic stroke induced network - both are identical
with the exception of reducing the diameter of a microvessel to simulate a ischaemic
stroke. A comparison in the spatial distribution of hypoxia and blood pressures between
the two networks will be made in order to try and understand the changes in network
environment caused by stroke.
A novelty of this report’s work is the use of brand new datasets. Structural data of
the mouse brain was acquired using a vascular casting technique with the resulting cast
scanned using Micro-CT. Current techniques can already extract blood vessels down to a
resolution of 5 µm, whereas brain capillaries can be as low as 3 µm [12]. Therefore, using
this method, an incomplete vascular network was collected - hence the use of the flow
estimation algorithm of Fry et al. [16] in the modelling. The modelling also makes use
of MRI fluid distribution data. A key motivation behind this report is that imaging can
only get flow information on a coarse scale, whereas the mathematical model can predict
fluid flow in individual vessel. Detailed information on oxygenation is also not currently
available using imaging - data which could prove vital in the treatment of stroke. Thus,
the modelling provides a tool to take the structural and fluid flow information to predict
blood flow and oxygenation.
The mathematical modelling process uses these data to parameterise the models and
also to compare against the model predictions of fluid distribution. A 1D Poiseuille flow
model, which describes a linear relationship between the flux and the pressure gradi-
ent along a vessel segment, is used, combined with in vivo blood rheology descriptions
such as the Fahræus-Lindqvist effect [34]. The oxygen transport model includes oxy-
gen bound to erythrocytes, dissolved within blood plasma, diffusion through the tissue
†Extract taken from Sweeney et al. [51].
13
and uptake by the cellular population [45]. The flows in major feeding and draining
vessels can be measured by MRI or estimated from literature values of brain perfusion;
this information parameterises boundary conditions for the larger vessels. Boundary
conditions for terminal micro vessels remain unknown but are required for a fully de-
termined linear system for the networks fluid pressures. An algorithm by Fry et al. [16]
which minimises the deviation of pressures and wall shear stresses from specified target
values, resolves this indeterminacy. Oxygen partial pressure levels are needed on the in-
flow boundary nodes in the oxygen transport model, these figures are based on current
literature. The resulting models and boundary conditions form a system of ordinary
differential equations representing conservation of mass and momentum, which are then
solved using numerical techniques in C++.
An ischaemic stroke is caused by a sudden blockage of blood flow to the central nervous
system. The most common causes are by thrombosis (a blood clot within a vessel) or
embolism (a foreign object such as plaque or a broken off section of a blood clot which
has been carried in the bloodstream). A loss of blood behind the stroke creates a hypoxic
environment which can lead to necrosis (cell death). The location of the ischaemic stroke
can have drastic effects, however, the size of the hypoxic region does not necessarily
correlate to the magnitude of the neurological damage. For example, a very small area
of hypoxia in the brainstem can be more devastating compared to relatively large damage
to the cerebellum or cerebral hemispheres [42]. Current therapies for stroke are designed
to restore blood flow to the hypoxic region as soon as possible after the blockage event.
However, even if the blockage is removed, the capillaries may be too damaged for efficient
perfusion to the surrounding tissue, or muscle cells may have undergone necrosis due to
a lack of oxygen. This is known as the no-reflow phenomenon. Hence, time is a crucial
factor when treating stroke.
In order to simulate an ischaemic stroke, a healthy simulation is taken and then an
individual microvessel is blocked. This is simply done by reducing the chosen microvessel
diameter to ∼ 0µm. A comparison in the spatial distribution of hypoxia and blood
pressures between the healthy and stroke induced networks will be made. With access to
incredibly detailed data sets on vascular structure collected through imaging techniques,
functional information can be extracted from these simulations which can be used for
stroke identification and treatment.
14
1.2 Brain Anatomy
(a) Superficial dissection of theright side of the neck, showingthe carotid and subclavian arter-ies.
(b) Diagram of the Circle ofWillis. This is regarded as thenormal configuration, however,fewer than half have this shape.
Figure 1.1: Reproducedfrom Gray [19].
The brain can be divided lengthwise down its middle
into what are known as the cerebral hemispheres. Each
hemisphere is subdivided into four lobes (frontal, oc-
cipital, parietal and temporal) by sulci and gryi (the
grooves and bumps on the surface of the brain). Within
the fluid-filled ventricles (the cavities between the lobes)
is the cerebrospinal fluid (CSF), whose function is to act
as a buffer, reduce pressure at the base of the brain, ex-
crete waste products and act as a endocrine medium for
the brain. Absorption of CSF into the blood stream can
occur in the superior sagittal sinus when CSF pressure is
greater than venous pressure, however, diffusion is only
unidirectional.
Metabolically, the brain is very active. A stable and
copious blood supply is require as the brain has no ef-
fective means of storing oxygen or glucose - the brain
accounts for 25% of the body’s oxygen consumption
[42]. The overall flow rate is maintained at a con-
stant level but may vary in certain regions depending
on neural activity. Currently the mechanisms control-
ling cerebral blood flow are not fully understood, but
there is general agreement that the following methods
are used. The first process is called autoregulation,
in which blood vessel themselves act to maintain con-
stant flow by constricting (therefore increasing flow re-
sistance) in response to increased blood pressure and
relax due to a decrease in pressure (decreasing flow re-
sistance). The second mechanism is a local response by
neurones, astrocytes and cerebral vessels to increased
neural activity.
1.2.1 Arterial cerebral circulation
Arterial cerebral circulation can be divided into two
categories, anterior and posterior cerebral circulation.
Anterior cerebral circulation is supplied by the internal
carotid artery (80% of supply) which bifurcates into the
15
anterior and middle cerebral arteries. Before bifurcation occurs, the internal carotid
artery gives rise to two smaller branches the anterior charoidal and posterior communi-
cating arteries. The posterior cerebral circulation is supplied by the vertebral arteries
(remaining 20% supply to the brain), which consists of the basilar, posterior cerebral
and posterior communicating arteries.
1.2.2 Circle of Willis
The Circle of Willis is an arterial polygon formed by the connection of the posterior
cerebral and the internal carotid arteries (see Figure 1.1b). Normally little blood flows
through the Circle of Willis since the fluid pressures in the internal carotid arteries are
approximately the same as that in the posterior cerebral arteries, leading to little blood
flow through the posterior communicating arteries. However, if a major vessel becomes
occluded over a sustained period of time, it can provide an alternative route to prevent
neurological damage. However, since the anterior and posterior communicating arteries
vary largely in size, in the case of an abrupt blockage, blood vessels my rupture.
1.2.3 Brain drainage
The principle route for venous drainage of the brain is through a system of cerebral
veins that empty into the dural venous sinuses and ultimately into the internal jugular
veins. This is often considered the only significant route through which venous blood
can leave the head, however, only in a supine position. Otherwise, the internal jugulars
collapse and blood leaves through the vertebral venous plexus instead.
Cerebral veins are conventionally divided into two group, superficial veins which lie on
the surface of the cerebral hemispheres (emptying into the superior sagittal sinus), and
(a) Superior sagittal sinus laid openafter removal of the skull cap. (b) Sagittal section of the skull.
Figure 1.2: Reproduced from Gray [19].
16
deep veins which drain internal structures (emptying into the straight sinus). Superficial
veins are quite variable and only three veins are reasonably constant from one brain
to another - the superficial middle cerebral vein, the superior anastomotic vein and
the inferior anastomotic vein. On the other hand, deep veins are more constant in
configuration.
Note, in addition to superficial and deep veins, there is a separate and complex col-
lection of veins that serves the cerebellum and brainstem.
1.2.4 Morphological features of intracortical arteries and veins
The vascular network of a brain contains complex and dense system of branch and
mesh-like blood vessels. Duvernoy et al. [12] categorised intracortical vessels into six
groups:
1. Group 1 - composed of vessels which reach the molecular layer (layer I) with the
possibility of extending into the external granular layer (layer II).
2. Group 2 - continues into the superficial part of the pyramidal layer (layer III a &
b).
3. Group 3 - are the most numerous vessels, which penetrate as far as the high
vascularised middle region of the cortex (centred in the internal granular layer IV
while overlapping in the pyramidal, III c and ganglionic, V a, layers).
Figure 1.3: Morphological features of intracortical arteries and veins. Arteries and veinsare divided into six groups, Ai and Vi where i = 1, . . . , 6, respectively. Reproduced fromDuvernoy et al. [12].
17
(a) General view of cortical vessels.(1) Cortical arteries. (2) Cortical vein.(3) Medullary artery (type A6). (4)Subcortical white matter. (b) Coiling of an artery of type A3.
Figure 1.4: Reproduced from Duvernoy et al. [12].
4. Group 4 - these vessels reach the multiform layer and inner limit of the subcortical
white matter.
5. Group 5 - vessels are comprised of both arteries and veins which pass through the
cortex and adjacent white matter.
6. Group 6 - a special group of large arteries which pass through the grey matter
without branching and supply blood only to the white matter.
Current literature conforms with morphological analysis in that vasculature takes
the form of tree (arteries/veins) and mesh (capillaries) like structures [26]. Chapter
2.2 uses these branching and looping formations to create an algorithm that analyses
a microvascular network’s topology and distinguishes between arterioles, venules and
capillaries. Our aim is to apply this algorithm to a dense brain network in order to
distinguish between arteries, veins and capillaries.
Work by Duvernoy et al. [12] only serves to show the complexities of aiming to simulate
blood perfusion in an entire brain network. Not only will the process be computationally
demanding due to the shear vascular density of brain networks (see Figure 1.4a) but
potentially challenging in terms of imaging due to various topological structures such as
coiling of arterial branches (see Figure 1.4b).
1.3 Hemodynamics
A vital component of modelling blood perfusion through vascular networks is using bio-
logically realistic mathematical descriptions of blood rheology. Throughout the human
18
Figure 1.5: Blood constituents. Reproduced and altered from [7].
body blood flow displays different characteristics, from inertia driven flow in the aorta
to the highly non-Newtonian, viscous driven flow at the capillary scale. Mathematical
modelling of these properties is currently on going. Presently, In vitro and in vivo mea-
surements are used to further our understanding of these mechanisms governing fluid
transport.
The following sections will discuss the fundamental biological factors in the study
of hemodynamics which become particularly important at the non-Newtonian capillary
scale. The mathematical descriptions will also be presented in order to implement them
in the discrete fluid transport model presented in Chapter 3.
1.3.1 Erythrocytes
Blood is a fluid vital in sustaining life in animals. The bodily fluid delivers integral
nutrients and oxygen to cells and transports metabolic waste products away from said
cells. Blood consists of erythrocytes (red blood cells or RBCs), thrombocytes (platelets),
leukocytes (white blood cells) and plasma. Blood plasma constitutes ∼ 55% of blood
volume and is essentially∼ 92% water, with the remaining∼ 8% being dissolved proteins
and trace amounts of other minerals. RBCs take up a much higher volume within the
blood when compared to other cells - normal human blood has a hematocrit (the volume
percentage of RBCs contained in blood - also known as packed cell volume or PVC) in
the range of 37− 48% for women and 42− 52% for men [4].
In vertebrates, RBCs are the principle means of delivery oxygen to tissue. Uptake
of oxygen happens within the lungs or gills and is released into surrounding tissue
while squeezing through capillaries. RBCs are biconcave disks approximately 8µm in
diameter. The disk structure is very flexible, so much so that when entering into very
thin capillaries, they deform into bullet-like shapes with approximately a minimum
19
Figure 1.6: Micrographs of RBCs flowing through small-bore glass tubes (those on theleft) and mesenteric microvessels (those on the right). In the vessel diameter range of∼ 6 − 13µm, cells may flow in single file at low hematocrits, HD. With increasinghematocrit, more cells are present in a vessel cross-section, leading to a transition tomulti-file flow and increased interaction of cells with the wall. Reproduced from Priesand Secomb [34].
diameter of 2.8µm [46]. Hence, the size of a RBC is relative to the capillary diameter
and has a significant impact on the apparent viscosity of the blood, and therefore, its
non-Newtonian characteristics. The non-Newtonian nature of blood is thought to be
important in vessels of diameters less than 300µm [37].
In large vessel or a high hematocrit, the interactions between cells, deformation of
said cells, combined with plasma and vessel wall interactions are all integral parts of
fluid transport. Whereas, in capillaries of less than 6µm, RBCs are forced into single-
file flow and so assumptions can be made, such as Lubrication Theory to model blood
plasma between the cells and cell wall, or assuming an axisymmetric erythrocyte shape.
Attempts have been made to model individual cells by assuming axisymmetric cells [46]
or by focusing attention to membrane deformation [13]. However, three-dimensional
simulations of more than a few cells is very computationally expensive and so currently
is not feasible. Alternatively, continuum descriptions can be used.
1.3.2 The Fahræus Effect
The Fahræus effect is the apparent decrease of hematocrit with decreasing capillary di-
ameter [14]. The velocity profile of Poiseuille flow (see Chapter 3 for details) is parabolic,
with no-slip boundary conditions at the walls (zero velocity condition) and has maximal
velocity down the vessel centre. Hence, due to the biological properties of blood, RBCs
travel along the centre of the vessel faster than the lubricating plasma layer adjacent to
20
the vessel walls.
A parametric description of the reduction of tube hematocrit, HT , relative to discharge
hematocrit, HD, captures the phenomenon of the Fahræus effect. The tube hematocrit
is the cross-sectionally averaged RBC concentration and is defined as
HT =1
A
∫θdA, (1.1)
where θ is the RBC concentration field and A is the capillary cross-section. The discharge
hematocrit is defined as the cross-sectionally averaged RBC concentration weighted by
blood velocity and thus defined as
HD =
∫vθdA∫vdA
, (1.2)
where v is the blood velocity field.
Based on experiments in which human RBC suspensions were perfused through glass
tubes with varied diameters, Pries et al. [38] derived an empirical relationship between
tube and discharge hematocrit. Mathematically the relationship can be described as
Structural data of the vasculature of the mouse brain was ob-
tained by Angela D’Esposito† through a novel method called vascu-
lar casting, a process which involves an invasive method of injecting
a casting agent into the vascular system of the mouse. The brain
was then extracted and scanned using Micro-CT. The resolution of
Micro-CT is 5 µm, given capillaries can be 3 µm in diameter, a full
dataset of the mouse brain is not expected. Analysis of vessel diam-
eters (see Figure 2.2a) shows a large number of diameters close to
the mean of 6.61 µm, but two diameters were found to be 2.5 µm,
suggesting a potential error the segmentation and skeletonisation
step.
The process of segmentation and skeletonisation of the resulting
dataset was done by Simon Walker-Samuel†. Segmentation of the
network allows for application of the discrete fluid transport model.
During this process the vascular network is sectioned into a series
(a) (b)
Figure 2.2: Plots of the frequency (logarithmically scaled) distributions of diametersand lengths (µm).
†Member of CABI, UCL.
26
(a) (b)(c)
Figure 2.3: 3D visualisation of the collected mouse brain network. (a) The full dataset.(b) & (c) the largest subnetwork. Vessels are to scale and colour is diameter dependent.
of vessel segments (sections of blood vessels that contain no bifurcations) which are
connected by by nodes (junctions between vessels - these internal nodes can connect
multiple segments) or in the case of a boundary vessel, a segment has a boundary node
(no connection to another segment at a boundary node i.e. a node connected to only
one segment) - see Figure 2.1. Frequency distributions of vessel diameters and lengths
can be seen in Figure 2.2. Diameters ranged from 2.5 to 67.4 µm with a mean of 6.61
µm. Whereas lengths ranged from 2 to 154 µm with a mean of 12.7 µm.
In order to parametrise the discrete fluid transport model in terms of boundary condi-
tions, phase contrast MRI was used to gain average blood flow velocities for the internal
carotid arteries (obtained by Rajiv Ramasawmy†). The boundary conditions are dis-
cussed in further detail in Section 3.3, but currently the mathematical modelling of these
networks is hindered by the lack of detailed flow information.
2.1 Largest Subnetwork Algorithm
Due to resolution limitations of Micro-CT and potential incomplete perfusion of the
casting agent around the mouse brain, analysis showed a scattering of disconnected
subnetworks. Perfusion occurred in arteries, some arterioles and very few capillaries,
therefore, none occurred in the venous side of the brain (confirmed by vessel diameters
and morphological comparisons). Hence, the next step was to design an algorithm to
extract the largest subnetwork (i.e. the subnetwork with the largest number of vessel
segments) in the dataset (see Algorithm 1) in order to have a coherent structure to
run discrete fluid flow simulations on. To reduce the run time of Algorithm 1 isolated
(single) vessels segments were removed from the dataset along with duplicate vessel
nodes (which connect vessel segments) - an artefact from segmentation.
The largest subnetwork algorithm starts by iterating through all boundary nodes. For
the ith boundary node, the segment connected to said node, Si is found and added to the
27
set of reached segments, RS . The algorithm then checks to see if Si has been connected
to a previously found subnetwork, if it has, the algorithm iterates to the next boundary
node. Otherwise, the convergence loop (see Algorithm 1) starts to iterate for a preset
number of iterations. Inside the loop, all daughter branches, Sk, of segments contained
within RS but are not members themselves are found and then added to RS . The loop
continues to find each generation of daughter branches until no vessel segments are being
added to RS - the algorithm has converged on the entire subnetwork. At this point the
subnetwork is stored in the matrix MI . Once all boundary nodes have been iterated
through, the largest subnetwork is then extracted from MI . MI itself is a matrix of
size nBC ×nS , where nBC and nS are the total number of network boundary nodes and
segments, respectively. The matrix acts as a ”flagging” system for identifying which
segments are assigned to the subnetwork associated with a boundary node. From here,
network data such as diameter and length distributions can easily be extracted.
Note, the largest subnetwork algorithm was tested on the mesentery networks of Pries
et al. [32]. For this report the largest subnetwork was wanted, however, this algorithm
can be easily adapted to find, say, the longest subnetwork or the subnetwork containing
the highest volume of blood.
2.2 Geometrical Analysis
When analysing the vascular network of a brain it is useful to be able to categorise the
vessels in terms of arterioles, venules and capillaries as in future work (see Discussion),
each will be treated using a different mathematical model. A novel geometry-based al-
gorithm by Smith [49] characterises the distinction between branching arterioles/venules
and mesh-like capillaries, the transition from the different topologies is marked by iden-
tifying loops within the network. The algorithm can be applied to an arbitrary network
nN nodes and nS segments, each with a given vessel diameter, Dj .
The pseudo code for the geometry-based algorithm is presented in Algorithm 2. The
algorithm starts by selecting the main feeding arteriole, which, if pressures are known, is
(a) (b)
Figure 2.4: Initial stages of the geometry-based algorithm applied to rat mesentery databy Pries et al. [32]. (a) the 1st and (b) the 3rd iteration. Figure has been reproducedfrom Smith [49].
28
Algorithm 1 Largest subnetwork algorithm.
Set iterMAX . . iterMAX - maximum number of iterations for convergence loop.Initialise MI = ∅. . MI - a matrix which stores a record of all subnetworks.for i = 1 : nBC do . nBC - number of boundary nodes.
Initialise RS = {Si}. . RS - the set of reached segments.. Si - segment who’s start node is the ith boundary node.
if Si /∈MI then . If Si has not been recorded in a previous subnetwork.ConvergeCounter = 0.for j = 1 : iterMAX do . Convergence loop.
Counter = 0.for k = 1 : nS do
if Sk is a daughter of segment in RS and Sk /∈ RS . thenAdd Sk to RS .
end ifend forSet Counter. . Counter - the number of segments in RS .if Counter = ConvergeCounter then
Terminate j loop.else
ConvergeCounter = Counter.end if
end forend ifStore subnetwork in MI .Set RS = ∅.
end forExtract largest subnetwork from MI .
the highest nodal pressure. This vessel segment is then classed as a parent and added to
the set of segments that have been classed as parent vessel, P , and the reached segment
set RS (likewise its corresponding nodes are added to the set of reached nodes, RN ).
On the ith iteration, parent segment pi is found to be the maximal diameter vessel
available (promoting the characteristic that arterioles or venules have larger diameters
when compared to capillaries) that is contained in RS but has not been previously classed
as a parent - pi and its corresponding end node is added to P and RN respectively. Its
corresponding daughter vessels, dj , are then found and added to RS .
The parent and daughter vessels are then classified by the following conditions:
1. If the end node, nj , of the daughter vessel, dj , is contained within the set of reached
nodes, RN , dj is a capillary.
2. If the above condition is satisfied and the diameter of the parent vessel, pi, is
smaller than the critical parent-daughter diameter ratio multiplied by the diameter
of the daughter vessel (i.e. Dr ×D(dj)), pi is a capillary.
29
3. If pi is classified as a capillary, then all of its daughter vessels are capillaries.
4. Else if the above conditions are not satisfied, pi is classified as an arteriole/venule.
The geometry-based algorithm by Smith [49] is run from both the arteriolar and
venular trees i.e. when classifying venules, the first parent segment is that which has
the lowest nodal pressure. The two results are compared, those vessels not classified as
either arterioles of venules are labelled as capillaries. The choice of Dr is a crucial factor
when running this algorithm as it can cause an overlap of the arteriole and venular trees,
which biologically is not correct as each tree has distinct pressure profiles. Smith [49]’s
choice of Dr was based on a comparison of results by Roy et al. [40] on rat mesentery
networks obtained by Pries et al. [32]. The method by Roy et al. [40] relies on predefined
flow directions to identify connected diverging or converging tree structures.
Work on implementing the geometry-based algorithm by Smith [49] is currently ongo-
ing. The mesentery network (from the Pries et al. [32] collection) is being used a testing
bed during implementation as it allows for a comparison with the work of Smith [49]
and Roy et al. [40]. The current result is shown in Figure 2.6c (Dr = 2.675), visually
it displays similarities to both Figures 2.6a and 2.6b. However, the algorithm currently
does not work when applied to the brain subnetwork extracted in Section 2.1.
The objective is to apply the geometry-based algorithm to our brain networks to
(a)
(b)
(c)
Figure 2.5: (a) Later stages of the geometry-based algorithm on the mesentery networkof Pries et al. [32]. (b) Magnification of the region outlined in (a). (c) the capillaryclassification of daughter branch, d2, one step later. Figure has been reproduced fromSmith [49].
30
study macro and microangiopathy (see Discussion for further details). Difficulty may
arise when selecting Dr (currently, sensitivity analysis for the brain networks has not
been done) for our networks and so the method by Roy et al. [40] or an alternative
geometry-based method by Cassot et al. [6] may need to be considered.
Algorithm 2 Geometry-based algorithm by Smith [49].
Select first parent segment p1. . p1 : the main feeding or draining vessel.Initialise P = ∅. . P : the set of segments which have been parents.Initialise RS = {p1}. . RS : the set of reached segments.Initialise RN = {start node of p1} ∪ {endnode of p1}.
. RN : the set of reached nodes.for i = 1 : nS do . Loop through parent segments.
pi ← maxD(s ∈ {RS\P}).. pi : the max diameter segment ∈ RS but /∈ P .
Add pi to P .Find daughter vessels dj (j = 1, . . . , nD) of parent pi.
. nD : the number of daughter vessels of pi.Add start/endnode of pi to RNAdd dj (j = 1, . . . , nD) to RS .for j = 1 : nD do
Find nj . . nj : end node of dj .if nj ∈ RN then
dj is a capillary.if D(pi) < Dr ×D(dj) then
. Dr : critical parent-daughter diameter ratio.pi is a capillary.
end ifelse
Add nj to RN .end if
end forif pi is a capillary then
dj (j = 1, . . . , nD) are also capillaries.else
pi is an arteriole/venule.end if
end for
31
(a) (b)
(c)
Figure 2.6: Geometrical analysis of the rat mesentery network. (a) Roy et al. [40].(b) Reproduced from Smith [49]. (c) Attempt at implementation of the geometry-basedalgorithm. Red - arterioles, green - capillaries and blue - venules.
Figure 2.7: Comparison of the number of arterioles, capillaries and venules presentedby Smith [49] and implementation for this report. Roy et al. [40] data not available.
32
3 Discrete Modelling of Microvascular
Fluid Transport
The objectives of this chapter is to introduce the concept and modelling of discrete
microvascular fluid transport. A basic model is presented which can be used when all
boundary conditions are known. An extension of this basic model is given, which uses
a flow estimation algorithm by Fry et al. [16] for the case when there are unknown
boundary conditions.
The key components when modelling blood flow within a discrete microvascular net-
work are network geometry and topology, blood rheology within vessels, in addition to
flow or pressure boundary conditions at all boundaries of the network. Blood flow in
microvascular networks is viscous dominated, along with the assumptions of Lubrica-
tion Theory the Navier-Stokes equations can be reduced. Application of Lubrication
Theory assumes blood is modelled as a continuum, and blood vessels are long and thin.
Mathematically, the assumptions can be defined as,
ε =D
L� 1 and ε2Re� 1, (3.1)
where ε is the aspect ratio between a vessel’s diameter, D, and length, L. Re is the
Reynolds number (a ratio between inertial and viscous forces) of the fluid. Under these
assumptions, Poiseuille’s Law, a linear relationship between a vessel (represented by an
axisymmetric tube) flux, Q, and its pressure gradient, can be used
Q = − πd4
128µ(d,HD)
dp
ds, (3.2)
where d is vessel diameter, µ is the dynamic viscosity, HD is blood discharge hematocrit,
p is the fluid pressure and s is the arc length down the central axis of the vessel.
A discrete microvascular network is a series of individual blood vessels, where Equa-
tion (3.2) can be applied to each individual vessel (along with network conservation
laws). The following section describes the process of modelling multiply connected ves-
sels combined with the previously seen in vivo model by Pries and Secomb [34], to
describe the flow resistance in microvessels.
33
3.1 Basic Model
As previously described in Chapter 2, to use a discrete model, the microvascular network
is segmented and represented by a series of vessel segments which are connected by
nodes and/or contain boundary nodes. For each vessel segment a positive flow direction
is defined from its start node to end node (this is arbitrarily chosen, although care needs
to be taken when defining flow at the boundaries i.e. positive or negative values for
inflow or outflow, respectively). Under the assumption of Poiseuille flow, the relationship
between the nodal pressures pk and the segment fluxes, Qj , can be expressed in the form
Qj =∑k∈N
Mjkpk, j ∈ S, (3.3)
where N and S are the set of all nodes and vessel segments in the microvascular network
respectively, and
Mjk =
+πd4j/ (128µjlj) , if k is the start node of segment j,
−πd4j/ (128µjlj) , if k is the end node of segment j,
0, otherwise,
(3.4)
where lj denotes the length of segment j.
As a consequence of the conservation of mass, the sum of the flow at each interior
node is zero. This can be combined with the boundary nodes to give∑j∈S
LijQj +Q0i = 0 for i ∈ N, (3.5)
where
Lij =
−1, if i is the start node of segment j,
+1, if i is the end node of segment j,
0, otherwise.
(3.6)
For all interior nodes, Q0i = 0. If i is a boundary node, Q0i is the inflow (or outflow
if negative). Combining (3.3) with (3.5) yields∑k∈N
Kikpk = −Q0i for i ∈ N, (3.7)
where
Kik =∑j∈S
LijMjk. (3.8)
Equation (3.7) forms a sparse linear system of nN (total number of nodes) equations.
If the pressure or flow is known at every boundary node (at least one pressure condi-
34
tion is needed), the system can be solved for nodal pressures using standard methods.
Substitution of resulting pressure solutions into equation (3.3) solves for segment fluxes.
3.2 Incorporating Variable Hematocrit
As detailed in Section 1.3.2, the modelling of variable hematocrit at microvascular bi-
furcations accounts for the non-Newtonian behaviour of blood perfusion at the capillary
scale. In this thesis, the empirical laws given by Pries et al. [38] are used to model this
process.
Absolute effective viscosity given by Equation (1.14), is a function of hematocrit and
since the fluxes, Qj , are needed to calculate vessel hematocrits, an iterative procedure
is used. To summarise the pseudo code presented in Algorithm 3, hematocrit is set to
an initial value (0.4) which is then used to calculated the absolute effective viscosity
on the first iteration. Equation (3.7) is then used to calculate vessel fluid pressures
and consequently segment fluxes from Equation (3.3). The nodes connecting vessel
segments are then listed in order of flow direction. From the list, a hematocrit boundary
condition is applied to a node if it is located at the network boundary. If the node
corresponds to a converging bifurcation, conservation of RBC flux is used. Finally, if
the node is a diverging bifurcation, the empirical law given by Equation (1.15) is used.
A relaxation factor (multiplied by a factor of 0.8 every fifth iteration) is included when
updating hematocrit values to ensure hematocrit convergence. Consequently, the new
and old values of vessel hematocrit and flux are compared and if the differences are
below previously defined tolerances, the algorithm terminates.
The algorithm for the phase separation iterative procedure is presented in Algorithm
3, however, implementation is ongoing and so constant hematocrit (with a value of 0.4)
is used for this report. Note, with constant hematocrit, Algorithm 3 terminates once
segment fluxes, Qj , are calculated.
3.3 Boundary Conditions
Rapid development for imaging of microvascular networks allows us to model detailed
three-dimensional networks. The distribution of blood flow rates fundamentally influ-
ences perfusion, solute transport, flow regulation, and growth and adaption in the mi-
crovascular networks but current available methods for observation of blood flow within
individual vessels is limiting. Experimental measurements of the main feeding arte-
riole(s) and draining venules(s) can be taken, but measurements of blood flow within
individual capillaries, in a capillary dense network such as the brain, would be extremely
time consuming, if not impossible using current methods. Thus, the mathematical mod-
35
Algorithm 3 Phase separation algorithm.
hOldj = hInit. . Defining initial hematocrit.relaxFactor = 1. . Set initial relaxation factor.hTol = 10−3 , qTol = 10−3. . Tolerances for hematocrit and flux.IterMax = 100. . Set the maximum number of iterations.for iter = 1 : iterMax do
if iter mod 5 ≡ 0 thenrelaxFactor ← 0.8× relaxFactor.
end ifCalculate µabs,j(dj , hOldj) using in vivo viscosity law - Equation (1.14).Solve Equation (3.7) for pressures, pk.From Equation (3.3), calculate fluxes, Qj .{nsort,inod} ← Generate list of nodes in order of flow direction.for iseg = 1 : nN do . nN - Total number of nodes.
if nsort,inod is a boundary node thenAssign hematocrit boundary condition.
else if nsort,inod is converging thenUse conservation of RBC flux.
else nsort,inod is divergingApply bifurcation law given by Equation (1.15).
. Assign initial value, τ0, for target shear stress, with randomly assigned signs.kp = 0.1.kτ = kτ,init. . Set initial value for kτ .Set kτ,final. . Assign final value for kτ .hTol = 10−3 , qTol = 10−3. . Tolerances for hematocrit and flux.while kτ ≤ kτ,final do
while flowSign changes dohOldj = hInit. . Defining initial hematocrit.relaxFactor = 1. . Set initial relaxation factor.for iter = 1 : iterMax do
if iter mod 5 ≡ 0 thenrelaxFactor ← 0.8× relaxFactor.
end ifCalculate µabs,j(dj , hOldj) using Equation (1.14).Solve Equation (3.7) for pressures, pk.From Equation (3.3), calculate fluxes, Qj .{nsort,inod} ← Generate list of nodes in order of flow direction.for iseg = 1 : nN do . nN - Total number of nodes.
if nsort,inod is a boundary node thenAssign hematocrit boundary condition.
else if nsort,inod is converging thenUse conservation of RBC flux.
else nsort,inod is divergingApply bifurcation law given by Equation (1.15).
Nonlinear iterations have converged. . Algorithm terminates.end if
end forflowSignj = sign(qj). . Record flow directions.τ0j = flowSignj × τ0. . Update shear stress signs.
end whilekτ ← 2× kτ .
end while
40
4 Analysis of Oxygen Transport to
Tissue
In this chapter, the governing equations describing oxygen transport are presented,
along with a Green’s function approach to this model. In this report oxygen transport
is simulated using the Green’s function method of Secomb et al. [45].
An essential function of the circulatory system is to delivery oxygen to tissue. The
distance that oxygen can diffuse into oxygen-consuming tissue is small, and so tissue
oxygenation is critically dependent on the spatial arrangement of microvessels in the
tissue. Thus, an ischeamic stroke at both the macro and micro scale can have a dras-
tic affect on oxygen consumption within tissue, eventually leading to the formation of
hypoxic regions within a circulatory network. Due to unavailable experimental infor-
mation, theoretical methods have been developed to simulate the spatial distribution of
oxygen levels in tissue surrounding a network of microvessels. Often blood is treated as
a continuum or computations are performed in the frame of reference of erythrocytes.
The first method struggles to capture the sharp O2 gradients near erythrocyte mem-
branes, and the second is problematic when assigning boundary conditions and so does
not allow for detailed information at the individual capillary scale.
Recent work by Lucker et al. [28] provided a new method to modelling oxygen trans-
port. Their model relies on moving meshes following individual erythrocytes, which are
then mapped over a fixed mesh serving for the calculation of oxygen transport in the
capillaries and surrounding tissue. The approach can account for variational hematocrit,
RBC velocity and metabolism. Lucker et al. [28] have run both steady-state and tran-
sient simulations to observe the influences of these parameters of PO2 (partial oxygen
pressure - the hypothetical pressure of O2 if it alone occupied the volume of the blood
at the same temperature) levels and have validated against PO2 measurements in the
rodent brain given by Parpaleix et al. [29]. Early simulations have been confined to a
single capillary network due to high computational costs, but work is currently being
done to expand simulations to larger networks.
Current work by Boas et al. [5] into capillary transit time heterogeneity, offers po-
tential when assigning PO2 boundary conditions. Boas et al. [5] made detailed mea-
surements of PO2 distribution from pail arterioles through a capillary network to pial
veins using a two photon microscope combined with an oxygen quenched phosphorescent
dye. This techniques allows to measure the speed and flux of individual RBC through
an individual vessel within a vascular network. Application of this method to vascular
41
networks of increasing size is time consuming, however, results could be used for average
PO2 levels in various regions of the brain. Boas et al. [5] also created a mathematical
model to calculate a PO2 distribution for a small subsection of the brain using a Krogh
cylinder model. The Krogh cylinder model, models solute transport from capillary to
surrounding tissues without the need of blood flow data. The structure of the model
implies that each capillary vessel is responsible for oxygen supply to the corresponding
surrounding cylindrical tissue section. However, there are multiple limitations to using
this model, for example, it assumes cylindrical angular symmetry around the central
axis while neglecting longitudinal diffusion of gas within the tissue, implying blood flow
is the sole mechanism for oxygen diffusion. The Krogh model also assumes there is
no diffusion of oxygen between surrounding tissue cylinders and as such, may not be
biologically realistic.
Finite-difference methods have been used for three-dimensional oxygen transport
problems by Goldman and Popel [17] & [18] and by Beard and Bassingthwaighte [2]
& [3], and for two-dimensional problems by Lo et al. [25]. For example, the following
equation is a two-dimensional time-dependent diffusion equation for the distribution of
PO2, P (x, y, t),∂P
∂t= Dα
(∂2P
∂x2+∂2P
∂y2
)− M0P
P + P0, (4.1)
where D is the diffusion coefficient of oxygen in tissue, α is the solubility, M0 is a
uniform oxygen demand in the tissue given a non-limiting supply of oxygen and P0
represents the PO2 at half-maximal consumption. A finite-difference method discretises
Equation (4.1) on to a grid and solves it using a numerical scheme. However, relative to
a Green’s function approach, the finite-difference method generally has a larger number
of unknowns in the numerical formulation and does not allow for rapid computations
for complex vascular geometries.
In this thesis, a Green’s function approach by Secomb et al. [45] is used to model
oxygen transport - a method derived from observations of skeletal muscle, brain, and
tumour tissues. In comparison to finite difference methods, the Green’s function ap-
proach reduces the number of unknowns in the numerical analysis and allows for faster
computations for complex networks.
4.1 Governing Equations
The following standard mathematical model describes oxygen transport through tissue
which is represented as a homogeneous medium with oxygen diffusivity D and solubility
α. Here, vessel walls are treated as part of tissue space and steady-state conditions are
42
assumed. Using Fick’s law of diffusion,
J = −D∇φ, (4.2)
which relates the diffusive flux, J , to the concentration, φ, combined with the conser-
vation of mass, leads to the governing equation
Dα∇2P = M(P ), (4.3)
where P is the tissue PO2 and M(P ) is the consumption rate. The dependence of
oxygen consumption on PO2 is represented by a Michaelis-Menten relationship
M(P ) =M0P
P0 + P, (4.4)
where M0 is a uniform oxygen demand in the tissue given a non-limiting supply of
oxygen and P0 represents the PO2 at half-maximal consumption.
The rate of convective oxygen transport along a vessel segment is given by
f(Pb) = Q [HDC0S(Pb) + αeffPb] , (4.5)
where Pb is the blood PO2 level (considered as an average value over the cross-section
of the blood vessel), Q is the blood flow rate, HD is the discharge hematocrit (the
volume flux of erythrocytes as a faction of the total volume flux in a vessel), C0 is the
concentration of haemoglobin-bound oxygen in a fully saturated erythrocyte, S is the
oxyhemoglobin saturation and αeff is the effective solubility of oxygen in blood. Both
Q and HD can be considered constant within a segment but may change from one vessel
segment to the next using the previously seen empirical bifurcation law given by Pries
et al. [38], whereas Pb varies with the position along a segment. The oxyhemoglobin
saturation, S, is represented by the Hill equation
S(Pb) =Pnb
Pnb + Pn50, (4.6)
where n is a constant and P50 is the PO2 at 50% saturation. The effective solubility of
oxygen in blood is given by
αeff = (1−HD)αp +HDαRBC , (4.7)
where αp and αRBC are the solubilities in blood plasma and RBCs, respectively. These
values are similar [22] and so αeff slightly depends on hematocrit and so can be approx-
imated by a constant value.
43
Conservation of oxygen infers
df(Pb)
ds= −qv(s) (4.8)
in each vessel segment, where s is variable defining segment length and qv is the rate
of diffusive oxygen efflux per unit of s.
Continuity of oxygen flux and PO2 must be enforced over the interface between blood
and tissue. Assuming a cylindrical segment leads to
qv(s) = −Dα∫ 2π
0
∂P
∂rrvdθ, (4.9)
where rv is the vessel radius and integration is performed over the circumference of
the vessel, denoted by the azimuthal angle θ. The blood vessel delivering the oxygen
generally exceeds local PO2 levels at the interface with the surrounding tissue. This
behaviour can be described by the following equation by Hellums [21],
Pv(s) = Pb(s)−Kqv(s), (4.10)
where Pv(s) is the average tissue PO2 and K represents intravascular resistance to
radial oxygen transport - a complex process, but here it is assumed K is a constant
dependent on each vessel’s diameter.
4.2 Green’s Function Approach to Oxygen Transport
In the following Green’s function method, blood vessels are represented as a set of
discrete oxygen sources. Using superposition principles, the resulting fields from these
sources represent the PO2 field in the tissue. As in equation (4.3), the tissue region is
represented by a set of discrete oxygen sinks. If the rate of oxygen uptake within the
tissue is known, the only unknowns in the problem are the strengths of both the sources
and sinks. The following mathematical model is presented by Secomb et al. [45] and is
used for this thesis.
Modelling oxygen transport, the Green’s function, G(x;x′), for a given tissue domain
may be defined as the PO2 at a point x resulting from a unit point source at x′. That
is to say, the Green’s function would be the solution to
Dα∇2G = −δ3(x− x′), (4.11)
44
where δ3 is the delta function in three-dimensions. The PO2 is thus given by
P (x) =
∫Sources
G(x;x′)q(x′)dx′, (4.12)
where q(x) represents the distribution of source strengths. The oxygen field resulting
down a long narrow blood vessel can be represented by a series of sources down its centre.
However, if it is assumed the distribution of these source strengths are equal to the
diffusive flux, qv(s), as defined by equation (4.9), the oxygen field close to the vessel may
not be accurately represented. In an attempt to improve this method, Hsu and Secomb
[23] assumed that the oxygen sources along the vessel axis must be uniformly distributed
around the circumference of the vessel on the blood-tissue boundary. However, Pozrikidis
and Farrow [30] noted that in order to satisfy a flux boundary condition, along with the
distribution of sources, a distribution of dipoles on the surface is generally required to
cancel the inwards flux into the blood vessel as a result of the distribution of sources.
Thus, implying the method of Hsu and Secomb [23] is an approximation. However, in
the case of oxygen transport, the continuous connectivity of vessel segments, gives rise
to smoothly varying oxygen sources and so the errors resulting from Hsu and Secomb
[23]’s approximation are small. Therefore, the approach of Hsu and Secomb [23] shall
be used in this approach.
4.2.1 Numerical Analysis
The following describes the numerical methods used to solve for oxygen transport using
the Green’s function approach. In a infinite domain, the solution is the singular function
G =1
4πDα|x− x′|. (4.13)
In a finite domain, additional nonsingular terms may be needed in order to satisfy
boundary conditions for the domain. If the microvascular network is segmented into nS
segments and the tissue is discretised into nT approximately cubic regions, a nS × nSmatrix of coefficients is defined in which Gvvij is the average tissue PO2 at the surface
of segment i as a result of a unit source distributed on the surface of segment j. Gvvij is
derived by integrating the Green’s function G(x;x′) with respect to both x and x′ over
the surface of segments i and j. An nS × nT matrix is formed using the coefficients of
Gvtij , which is defined as the tissue PO2 at the midpoint of segment i resulting from a
unit source at the midpoint of tissue region j. Likewise, a nT × nS is formed from the
coefficients of Gtvij , which is defined as the tissue PO2 at the midpoint of tissue region i
resulting from a unit source at the midpoint of segment j.
Use of the above definitions combined with equation (4.12), the average PO2 at both
45
the blood tissue interface of segment i and the tissue region i are given respectively by
Pv,i =
nS∑j=1
Gvvij qj +
nT∑j=1
Gvtijφj +G0, i ∈ S, (4.14a)
Pt,i =
nS∑j=1
Gtvij qj +
nT∑j=1
Gttijφj +G0, i ∈ T, (4.14b)
where qj is the source strength of segment j, φj is the sink strength of tissue region j
and T is the set of tissue points.
From Equation (4.10) we have
Pv,i = Pb,i −Kqi, i ∈ S, (4.15)
where Pb,i is the PO2 of blood at the midpoint of segment i.
From Equation (4.4) we have
φj = −M(Pt,j) = − M0Pt,jP0 + Pt,j
, i ∈ T, (4.16)
where Pt,j represents the PO2 in tissue region j as given by (4.14b). When calculating
the tissue PO2, in the case of no-flux boundary conditions on the tissue domain, or if
no boundary conditions are imposed, the constant G0 is required. G0 is determined by
the following conditionns∑j=1
qj +
nT∑j=1
φj = 0, (4.17)
which states that the net oxygen delivery is equal to oxygen consumption.
By integrating Equation (4.8) with respect to segment length, the rate of convective
oxygen transport along vessel segment i at its midpoint, f(Pbi), can be expressed as a
linear function of the source strengths qj ,
f(Pb,i) = f(P 0b,i)−
ns∑j=1
αijqj for i ∈ S; (4.18)
where P 0bi is the intravascular PO2 of segment i in the absence of diffusive oxygen
exchange. In this case, qj = 0 for all j and P 0bi is typically given by an assumed PO2
by an upstream vessel entering the tissue domain. For a series of segments forming an
unbranched vessel, αij is equal to 1 if segment j is upstream of segment i, 12 if i = j and
0 otherwise. If the segments form a branching network, αij depends on the partition of
oxygen fluxes at converging bifurcations.
46
Combining Equations (4.14a), (4.15) and (4.18) we get the following equation
ns∑j=1
Gvvij qj +G0 − f−1f(P 0
b,i)−nS∑j=1
αijqj
+Kqi
= −nT∑j=1
Gvtijφj , i ∈ S,
(4.19)
which is a set of ns equations to be solved together with Equation (4.17) for qj and G0.
C++ code generated by Secomb [43] is used to simulate the described Green’s function
model. The inputted network file is contained with a cuboid which must include all
oxygen-consuming regions. The midpoint of each vessel segment represents a oxygen
source, and so a segments efflux is a product of source strength and vessel length. The
tissue domain is split into subregions each of which are centred on a tissue node point.
In each of these subregions, the oxygen consumption rate is uniform and dependent on
the PO2 at the nodal point according to the Michaelis-Menten kinetics (Equation 4.4).
Next, inside the cuboid, the smallest convex region containing all tissue nodes within a
pre-defined distance of the nearest vessel is calculated. This convex region is equivalent
to the simulated tissue domain.
4.2.2 Boundary Conditions
In general, the boundary conditions on the exterior surface of the tissue domain cannot
be determined by available experimental information, in which case, additional assump-
tions must be made to form a well-posed problem. Two choices are imposing no-flux
(pointwise no-flux or net no-flux) or periodic boundary conditions. Imposing a no-flux
boundary condition is not realistic when applied to vascular networks, which tend to
have irregular geometries and therefore local oxygen fluxes can occur across any ar-
bitrarily chosen boundary in the tissue. Use of this boundary condition can result in
exaggeration of the amount of hypoxia in the tissue. However, using the Green’s function
method, the tissue volume can be considered as being embedded in an infinite domain
with the same diffusivity, where no oxygen sources or sinks are located externally to
the given tissue region. Hence, a well-posed problem is formed without imposing ex-
plicit boundary conditions on the outer surface of the tissue region. When compared
to finite-difference methods, the approach minimises artefacts associated with assigning
boundary conditions.
47
5 Mathematical Simulation of Fluid &
Mass Transport
In Chapter 3, the mathematics behind discrete fluid transport in blood vessel and blood
flow estimation of Fry et al. [16] was shown. In this Chapter, simulations of discrete
fluid transport on a brain subnetwork are run for both healthy and ischaemic stroke
induced networks. The subsequent flow solutions are then use in order to partially
parametrise the Green’s function approach to an oxygen transport model. The run-time
of simulations using these models greatly increases with respect to the number of vessel
segments contained within a network. Due to time limitations, a further network from
the largest subnetwork (see Figure 2.3b) was acquired.
Since averaged blood flow velocity data of the left and right carotid arteries was
obtained by MRI, the conditions for Lubrication Theory (see Equation 3.1) at the carotid
vessels where tested. For example, for one of the carotid arteries, the aspect ratio
was calculated as ε ≈ 0.07997 (where the diameter and length of the vessel are given
by 22.29µm and 278.747µm respectively) and the Reynolds number as Re ≈ 0.6045
(where the blood velocity, viscosity and density are given by 12.6 cm/s, 4.879 cP [52]
and 1050×10−18 kgµm−3 [24] respectively). Thus, the conditions for Lubrication Theory,
ε ≈ 0.07997 � 1 and ε2Re ≈ 0.0028 � 1, were thought to be within tolerances. Since
at the internal carotid arteries diameters are relatively large and the blood velocities
are the highest in the entire network, the discrete blood flow model is applicable to
the entire mouse brain data. Subsequently, a network was extracted in the vicinity of a
carotid artery. The network was retrieved by simply selecting a blood vessel and running
a simple algorithm to keep all vessels contained within a predefined box (4000× 5000×4000µm). The largest subnetwork algorithm was then run again to obtain the largest
network within the acquired data (see Figure 5.1b).
The region of the mouse brain in Figure 5.1b was selected due to the characteristics
of the Circle of Willis. A network with a looping structure was thought to possess more
desirable results when simulating an ischaemic stroke. For example, if a network was
chosen that has two branching regions connected by a a single blood vessel, simulating
an ischaemic stroke in this vessel could (dependent on boundary conditions) result in
an entire region of the network having zero blood flow. Comparatively, if an ischaemic
stroke occurred in a looping structure, more interesting results would be produced in
terms of how the network compensates blood flow through the remaining unblocked
vessel(s).
48
(a) (b) (c)
Figure 5.1: (a) The largest subnetwork obtained from the original mouse data. (b)Extracted section of the Circle of Willis, with the carotid artery circled. (c) Furtherreductions for the oxygen transport model. Colouring of vessels is diameter dependent.
The run-time of the oxygen transport solver varies drastically depending on the num-
ber of vessel segments within a network. Due to time limitations, reductions to the
number of vessel segments within the network used for this model had to be made (see
Figure 5.1c). To give a understanding of the sheer quantity of blood vessel segments
within the network and subsequent subnetworks - the full brain dataset shown in Fig-
ure 2.3a contains 186, 245 segments, the largest subnetwork (see Figure 5.1a) contains
72, 429 segments, the section of the Circle of Willis (see Figure 5.1b) contains 4, 718
segments and its reduction (see Figure 5.1c) only contains 515 segments.
5.1 Flow Estimation Method of Fry et al.
In this section, discrete flow solutions are derived for the sub-section in Figure 5.1b.
Limited physiological data on flow and pressure distributions in the vascular networks
are available, meaning the determining boundary conditions remains problematic when
simulating flows in discrete microvascular networks.
In this thesis, discrete flow solutions were obtained via the flow estimation method of
Fry et al. [16] introduced in Chapter 3. The optimisation procedure was developed and
tested on the mesentery networks of Pries et al. [32]. Parameter values were assigned
for the target pressure, the target shear stress and weighting of shear stress to pressure
terms. Based on diastolic pressures given by Czosnyka [8], target pressure was set at a
constant value of p0 = 40 mmHg. No data was found to allow accurate assignment of
a target shear stress value. Fry et al. [16] assumed the magnitude of the target shear
stress to be |τ0j | = 15 dyn/cm2, the average shear stress corresponding to the median
49
segment pressure across all four mesentery networks - to avoid obscure results, the same
constant value was used.
As per Fry et al. [16], kp was arbitrarily set to 0.1. The same value as Fry et al. [16]
was also used for kτ = 2.147×105. Uniform hematocrit boundary conditions of 0.4 were
applied at all inflow boundary nodes. Two pressure boundary conditions, based on data
by Czosnyka [8], were set to 60 mmHg on two vessels located on the Circle of Willis
(see Figure 5.2a). Using the MRI data collected for the internal carotid arteries, a flow
boundary condition at the carotid vessel was set to 12.6 cm/s.
5.1.1 Results
(a) (b)
Figure 5.2: 2D representations of the discrete flow solutions for pressures (mmHg). (a)The healthy network - flow boundary condition given by the solid circle and the twopressure conditions shown by the dashed circles. (b) Ischaemic stroke induced network- blocked vessel has been circled.
Blood fluid pressure values for both the healthy and ischaemic stroke induced networks
ranged from 48.16 to 66.85 mmHg (consistent with physiological results by Czosnyka [8])
and 45.86 to 61.01 mmHg respectively. Inducing an ischaemic stroke reduced the mean
pressure from 54.75 mmHg in the healthy network to 51.71 mmHg. Standard deviation
increased from 3.61 to 3.87 mmHg, indicating a slight increase in the range of pressure
values for the stroke induced network. This shift can easily be seen in Figure 5.3, where
blood vessel pressures fluctuate from 55 to 57 mmHg, towards the 46 to 49 mmHg range.
Visually this can be seen when comparing Figures 5.2a and 5.2b - the region north of
50
Figure 5.3: A plot comparing network pressures against the % of blood vessels withina network. The distributions of the healthy and stroke induced networks are overlaidfor ease of comparison.
the stroke experiences a pressure drop of ∼ 10 mmHg suggesting an increased likelihood
of hypoxia in this region.
5.2 Green’s Function Method for Oxygen Delivery
In this section, the Green’s function method see in Chapter 4 is used to simulated
oxygen transport from the extracted subnetwork of blood vessels to the finite volume of
tissue representing a portion of the mouse brain. As previously described, the numerical
method used treated vessels as a distribution of oxygen sources and the tissue as a
distribution of oxygen sinks.
As previously described, the discrete flow solutions parametrise the Green’s function
model. Since the number of vessel segments were reduced (see Figures 5.1b & 5.1c)
in our chosen network, the flow solutions are effected. Thus, the discrete flow model
was run again on the simpler network shown in Figure 5.1c. The resulting flow solu-
tions combined with hematocrit values for all vessel segments and the PO2 and solute
concentrations of boundary nodes (inflow nodes), were used to parametrise the Green’s
function model. Solute transport and reaction parameters are predefined as per the
C + + code provided by Secomb [43] - oxygen is handled differently compared to other
solutes to account for binding by haemoglobin. Intravascular resistance to oxygen is also
parameterised as per Secomb [43].
The cuboidal box containing the network was reduce as compactly as possible - the
size was reduced to 3580×4430×3910µm. The solver subdivides this domain, with each
subdivision centred on a tissue node point. The simulations were parametrised so that
51
there were 50 tissue points in all dimensional directions. The algorithm approximates
the smallest convex region inside the cuboid in order to simulate the tissue domain. The
smallest convex region includes all tissue node points within a distance of the nearest
vessel - this distance was specified as 30µm.
5.2.1 Results
Comparing Figures 5.4a and 5.4b, it is clear the ischaemic stroke induces oxygen depri-
vation in both the surrounding blood vessels and brain tissue in the chosen PO2 planar
slice. Inducing an ischaemic stroke reduced the mean tissue PO2 level from 33.76 mmHg
to 13.5 mmHg, while decreasing the standard deviation from 39.92 to 33.38 mmHg -
indicating a shift of PO2 distributions more towards the mean in the stroke network.
This distribution shift is apparent in Figure 5.5b, suggesting an increase in potential
oxygen deprivation. Note, in Figure 5.5a there is a significant increase (∼ 17%) in tis-
sue with a PO2 of zero, implying a large increase in hypoxia within the stroke induced
brain tissue. This would pose a problem to administering drugs to the patient as drug
transport would be affected in the same way as oxygen transport.
(a) (b)
Figure 5.4: A 2D planar slice representing PO2 (mmHg) in the blood vessels andsurrounding tissue for both the (a) healthy and (b) stroke (circled) induced network.
52
(a) (b)
Figure 5.5: Frequency distributions for PO2 levels in the surrounding tissue for boththe healthy and stroke networks. (b) is a zoomed in region for the 1 to 40 mmHg ofPO2 range of (a).
53
6 Discussion
The brain demands high quantities of oxygen which is distributed through its densely
packed vascular network. Under normal conditions, steep gradients of PO2 are found
in the vicinity of microvessels but low levels may be found in tissue [44]. Hence, any
decrease in perfusion as a result of a stroke can lead to tissue hypoxia and potential
necrosis - affecting the treatment and recovery of a patient suffering from stroke.
This report has presented a combination of mathematical models which simulate blood
perfusion and oxygenation in a microvascular network. These models were applied to a
network containing a section of the Circle of Willis. Simulations were run to model fluid
transport in a healthy and ischaemic stroke induced network in order to give us a insight
into the mechanics behind blood and O2 transport. Work focused on an ischaemic stroke
occurring in just one vessel, but future study would aim to understand the variability
of stroke in different areas of the brain.
Chapter 1’s main focus was introducing current mathematical models which describe
non-Newtonian blood rheology phenomena. The bifurcation law of Pries and Secomb
[34] attempts to resolve the disparity between in vitro and in vivo data by combining
in vitro blood viscosity effects with effects of a diameter dependent endothelial surface
layer. However, the cause of the discrepancy remains to be fully known and could be a
combination of factors, including ESL. A limitation of this model is that it was developed
for bifurcations with three vessel segments connected to one node, whereas within the
capillary network of a brain there are often multiple vessels attached to a node. The
empirical phase separation model of Pries et al. [31] was also presented. Future work
following on from this report would be to complete the integration of this empirical
formula into the discrete fluid transport model. However, work by Guibert et al. [20],
who simulated cerebral blood flow in primate cortex, found that phase separation had
little effect on network-wide hemodynamic properties when imposing uniform hematocrit
boundary conditions. Thus, neglecting phase separation in a large brain network may
be a reasonable assumption, but obviously more work is needed.
Chapter 2 presented a geometry-based algorithm by Smith [49]. During the process of
running fluid transport simulations on the subnetwork of the Circle of Willis, it became
clear that the algorithm of Smith [49] would not work on an entire brain network. The
geometry-based algorithm uses the topological structures of arteriolar/venular trees and
capillary meshes to classify blood vessels. It was hoped that this could be used on a brain
network to distinguish between arteries, capillaries and veins. However, considering the
Circle of Willis is a looping structure composed of arteries, the algorithm by Smith
54
[49] would classify it as a capillary network. Therefore, a modification of the algorithm
or use of perhaps the flow-based algorithm by Roy et al. [40] may be preferential for
application to an entire brain network.
The application of a algorithm to classify the blood vessels within a brain network
is not just interesting from a statistical point of view, it can be used as a tool into the
study of angiopathy. Angiopathy is a generic term for a disease of the blood vessels,
which comes in two forms, macroangiopathy and microangiopathy. Macroangiopathy is
the occurrence of atherosclerosis (the thickening of an artery wall as a result of invasion
and accumulation of white blood cells) - the result being the formation of a blood clot
preventing the flow of blood i.e. stroke. Microangiopathy is where the walls of small
blood vessels become so thick and weak that they bleed and leak protein, increasing
flow resistance and therefore slowing the flow of blood through the body. In this case,
the discrete fluid transport would be need to be modified to include Darcy’s Law (a
constitutive equation that describes the flow of a fluid through a porous medium), how-
ever, this will be discussed later. Application of a blood vessel classification algorithm
could lead to automatic simulation of either macro or microangiopathy (or selective use
of the discrete or continuum fluid transport models - discussed later) in the appropriate
vessels.
Chapter 3 presented the formulation of a discrete fluid transport model, along with
an adaption which incorporates the flow estimation algorithm of Fry et al. [16]. The al-
gorithm attempts to address the lack of experimental microcirculatory flow distribution
information by minimising the deviation from target wall shear stresses and pressure
values - a useful method when not all flow or pressure conditions are known at a net-
work’s boundaries. However, this approach has significant limitations as it is trying
to solve a system that is mathematically underdetermined when constrained by avail-
able data. Inevitably results are an estimate, with the number of known boundary
conditions greatly affecting the accuracy of the solution [16]. With increasing known
boundary conditions, the flow estimation errors are reduced. Considering the network
that was used for this report (see Figure 5.1b), only 3% of boundary conditions were
”known”, suggests the potential for a large amount of flow errors. Further investigation
is needed in order to derive the optimal balance of shear stress and pressure terms in the
optimisation procedure for a brain network as these parameters have a large weighting
on the discrete flow solutions. Nonetheless, Fry et al. [16]’s method could lead to new
insights into functional properties of microcirculation.
As mentioned in Chapter 2, vascular casting of a mouse brain is a complicated proce-
dure that can result in incomplete perfusion of the casting agent, and therefore retrieval
of a system of disconnected subnetworks that form a part of the mouse brain. Recent
work by Dobosz et al. [10] looks to overcome the limitations of conventional imag-
55
Figure 6.1: Visualisation of tumour morphology and tumour vascularisation. Scalebar, 250 µm and 100 µm (blowup). Reproduced from Dobosz et al. [10].
ing techniques. Dobosz et al. [10] applied multispectral fluorescence (UM) to the field
of tumour analysis and thus was able to create unprecedented three-dimensional and
quantitative insights into whole tumours with cellular resolution (see Figure 6.1). This
technique provides an exciting opportunity to procure an entire vascular dataset of a
mouse brain. The blood to the brain is supplied by the carotid and vertebral arteries,
and drained through the cerebral veins. Flow information for these blood vessels can
be acquired through MRI. Hence, with a full dataset, all boundary conditions would be
known and thus the flow estimation algorithm of Fry et al. [16] would not be needed in
the discrete model. In turn, this would lead to a significant reduction in the run-time
of simulations of discrete fluid transport, although considering the potential number of
vessel segments in such a dense dataset, the run-time would still be large.
Recent developments in mathematical modelling of microvascular networks presents
an exciting option for modelling blood flow in large incomplete networks. Shipley and
Chapman [47] developed continuum models for fluid transport through leaky neovas-
culature and porous interstitium of a solid tumour using mathematical homogenisation
methods. The multi-scale method exploits the separation of length scales between indi-
vidual capillaries and the tissue by assuming a periodic microstructure. This continuum
model has already been applied by Smith et al. [50] on coronary microcirculation. Fur-
ther development by Shipley et al. [48] has resulted in a coupled discrete-continuum
model for microcirculatory blood flow. Here, Poiseuille’s law is explicitly used to de-
scribe fluid transport through arteriolar vessels and coupled via point sources of flux
to a Darcy model for transport in the capillary bed. Considering current drawbacks of
56
network acquisition, the density of a brain network and limitations of the flow estima-
tions of Fry et al. [16], the discrete-continuum model provides an exciting option for the
mathematical modelling of fluid transport in large portions of the mouse brain.
Chapter 4 introduced the concept of a Green’s function approach by Secomb et al. [45]
to the transport of oxygen through blood vessels and into the surrounding tissue. The
novel approach allowed for faster computation of PO2 levels compared to finite difference
methods due to the lower number of unknowns. However, there is no experimental
data available giving detailed data of PO2 distributions in the mouse brain. Therefore,
the accuracy of this model of the microvascular networks of a brain cannot be fully
confirmed. Simulations of this model were run using external software provided by
Secomb [43]. Further work needs to be done to further understand the C + + coding
of this package to allow for a accurate adaption to the mouse brain, combined with an
increase in computational power to allow for simulations on more detailed networks.
57
7 Conclusion
This report has presented a novel combination of mathematical models for fluid and
mass transport in microvascular networks with applications to stroke research. Using
a discrete fluid transport model, experimental limitations are overcome by allowing for
detailed flow information at the individual blood vessel scale. The resulting information
was used to gain an insight into oxygen transport around the brain using a innovative
Green’s function model which allowed for a comprehensive study of oxygen diffusion at
the micro-scale. Understanding the processes of fluid and mass transport that occur
during an ischaemic stroke, could allow for the development of more effective methods
of treatment and improve our fundamental understanding of the disease state. Although
confirmation of these results are hard to come by due to experimental drawbacks, the
merging of these two models allows for insight into the deeply complex and relatively
unknown mechanisms occurring in the brain during this devastating illness. Recent
and future developments in both medical imaging and mathematical modelling offer
extremely exciting opportunities to build upon this report’s work, and to further our
understanding into the brain’s vascular environment during stroke, which can hopefully
help in the treatment and recovery of stroke patients.