Skip to content

Branched vessels ​

A vessel is a tree of branches. Every branch is a tube exactly as described in the previous pages: a centerline, its sections and frame, and a wall. The sections of most branches are star-shaped about their centerline and parametrized by their polar angle θ: the wall has no tangential component, and its radial one is the wall radius function ρb(τ,θ). A branch whose sections are far from star-shaped (an aneurysmal sac, a crescent) is parametrized by its own harmonic angle, as a tube is (Branches that are not star-shaped). An unbranched vessel is a tree of one branch, and for it everything below reduces to the tube encoder. Where branches meet, at junctions, the lumen of the vessel is the smooth union of the lumens of its branches.

Loading the model…
Drag to rotate, scroll or pinch to zoom, right-drag to pan

An aortic arch with its three supra-aortic branches (data/vmr/aortic-arch/0070_H_AO_H): the wall of the model coloured by branch, the centerline of every branch and the apex of every junction. The aorta continues past its three junctions; each branch is a tube.

The model ​

Each branch as an implicit tube ​

With (τb,θb,ρb)(x) the vessel coordinates of a point x in branch b, and rb the mean radius of the branch, the function

fb(x)=ρb(x)−ρb(τb(x),θb(x))rb

is negative inside the tube of branch b, zero on its wall and positive outside, in units of its radius. Its wall {fb=0} is exactly the surface of the tube model. For a branch parametrized by its harmonic angle, a ray from the centerline may meet the wall more than once, and fb is instead the signed distance from x to the nearest point of the wall in its section, db(x)/rb (negative inside, see Wall coordinates), which has the same zero set. Near the wall the two agree to first order.

Junctions and presence ​

At a junction, children leave a parent. Either the parent continues past it (a trunk from which side branches leave, such as the aorta and its supra-aortic branches) or it ends there and its children start (a bifurcation). The junction has an apex: the flow divider, or carina, in the crotch where the branches part.

Every branch is present along a range of its own τ through a smooth step

hb(τ)∈[0,1],

which is one along the branch and, at a junction where it starts (or ends), rises from zero to one (or falls) over a window that ends at the apex. The window is junctions.window parent radii long. A child is therefore present a little way back into its parent, and a parent that divides is present a little way past the apex, so that near a junction the branches overlap. The step is t3(6t2−15t+10), which is twice continuously differentiable.

The smooth union ​

The lumen of the vessel is the smooth union of the branch lumens:

F(x)=−εlog⁡∑bhb(τb(x))e−fb(x)/ε,wall={x:F(x)=0}.

Where one branch alone is present (hb=1 and the other terms vanish), F=fb: away from junctions every branch is exactly its tube. Where several overlap, F is a smooth minimum of their functions, so the lumen is their union. The parameter ε (junctions.smoothness, relative to each branch's radius) rounds the creases where two branch walls meet on a scale εrb, and the presence enters as −εlog⁡hb, so a branch fades out by shrinking. The weights

wb=∂F∂fb=hbe−fb/ε∑chce−fc/ε,∑bwb=1,

say how much each branch shapes the wall at a point. The branch with the largest weight is the one whose coordinates a point is given.

Finding the tree ​

Where two outlets part ​

Let φab be the harmonic function on the wall that is zero on the inlet, one on two outlets a and b, and has no flux through all the other outlets (they are dead ends). As φab decreases from one, the components of its superlevel sets {φab≥s} around a and around b grow towards the inlet and meet at a single saddle, in the crotch where the paths from the inlet to a and to b part. The merges of superlevel components are the split tree of the field (Carr et al. [24]), computed by sweeping the vertices in decreasing value with a union–find structure. Harmonic functions have been used before to find the skeleton of branched vessels (Zhu et al. [18]); here the pair fields give a definition of where two lineages part that needs no centerline and no choice of tolerance.

Clustering ​

The tree is built by single-linkage clustering (Sibson [25]): the two outlets whose lineages part farthest from the inlet, measured by the distance along the wall to the saddle, are joined first, and every join is a junction. A junction is described by the pair of its main lineages, the widest outlet on either side. Its saddle is the junction's apex, and its field says which side of the junction every vertex is on. This choice matters when side branches leave a trunk close together: two neighbouring side outlets at potential one lift the field on the wall between their mouths, and the saddle of that pair lies there instead of at a carina.

Which branch a point belongs to ​

A point belongs to the branches whose flow passes by it. With uo the harmonic field that is zero on the inlet and one on outlet o alone, |∇uo| divided by the flux ∫|∇uo|2 is the density of the flow to o: about one over the perimeter along the lineage of o, and decaying quickly in the branches that leave it. Following the tree from the inlet, a point goes past a junction to the side of one child when the strongest flow to the other child's outlets is less than half the strongest flow to that child's own, and stays before the junction when both pass by it.

Loading the model…
Drag to rotate, scroll or pinch to zoom, right-drag to pan

The synthetic bifurcation (furca synth bifurcation): the parent ends at the junction and its two children start. The apex is the saddle of the field of the two outlets, in the crotch.

Trunks and bifurcations ​

Consecutive stretches of the tree form one branch when the widest child continues its parent. The width of a stretch is the mean radius of its sections (below); a stretch too short to have sections, between two close junctions, takes the width of its widest continuation. When the widest child is junctions.continuation times wider than every other child (1.2 by default), it continues the parent: an aorta continues past its side branches. Otherwise the junction could be either, and the data decide: the vessel is fitted both ways, with the parent ending at the junction and all its children starting there (a bifurcation), and with the parent continuing into its widest child (a trunk). The trunk is chosen when its RMS residual is at most junctions.trunk_residual (one half) of the bifurcation's. Below one, the bifurcation, which treats its children alike, is kept unless the trunk fits clearly better: a symmetric bifurcation, such as that of the iliac arteries, fits about as well either way and divides its parent. A hypoplastic aortic arch, hardly wider than its brachiocephalic trunk, is a trunk: as a bifurcation, the backward extensions of its children leave a bump of most of the radius in the crotch where the arch curves away, and the trunk fits it with less than half the RMS residual (Validation).

Sections of a branch ​

As for a tube, the sections of a branch are the level rings of a harmonic field: its own field, zero on all the wall upstream of the branch and one beyond its end (its outlet, or the branches it divides into), harmonic on the branch and on the branches that leave it, which are dead ends there. Its rings are taken on the branch alone; a ring that crosses the mouth of a side branch is cut short and left out (it does not close). Near the start of a branch the rings follow the outline of its mouth, where the field is zero, so they are complete sections that include the flare of the mouth. Where a branch divides, its last rings pinch at the apex and wind around both children; those whose plane passes closer to the apex than junctions.clearance ring radii are left out. The centerline and the section tilt of each branch are fitted to its rings exactly as those of a tube (Centerline, Vessel coordinates).

The angular origin of a child is carried over from its parent: u(0) is the axis u of the parent's rotation-minimizing frame at the apex, rotated by the smallest rotation that takes the parent's tangent to the child's. The frame continues through the junction as it does along a bend, so that θ has the same meaning along the whole tree, and it is well defined for every direction a child leaves in (projecting the parent's axis onto the child's section plane would not be for a child leaving along that axis).

The walls ​

Domains ​

The wall of a branch is fitted where the branch has data of its own: to its free ends as for a tube, and at a junction as far as its sections are covered, where at least junctions.coverage of the angular sectors around it hold its own vertices. Near a junction the data of a branch end along the wavy outline of its mouth, and sections covered in only a few directions would leave its wall without support in the others. Beyond its domain, where it is still present, a branch continues as its end section along the straight extension of its centerline, for at most junctions.extension of its radii.

Over regions without data, such as the mouth of a side branch in the wall of its trunk, a penalized spline continues biharmonically: it carries the slopes at the rim into the hole and can overshoot. The coefficients without data (whose weight in the normal equations is below a twentieth of the mean) are therefore replaced by the harmonic interpolation of the coefficients around them, on the coefficient grid spaced as in physical space. A harmonic function stays between its values on the rim (maximum principle).

The joint fit ​

All walls are fitted to the input by minimizing ∑iAiF(xi)2 over the input vertices xi (with their areas Ai), plus the penalty of each wall, by block Gauss–Newton. One branch at a time is refitted with the others fixed. Since ∂F/∂ρb=−wb/rb, the linearized problem for a branch b parametrized by its polar angle is an ordinary wall fit of the targets

ρb(τb(xi),θb(xi))+F(xi)rbwb(xi)with weightsAiwb(xi)2,

over the vertices where it contributes at least a thousandth of the union. Where branch b is alone, the target is the vertex's own ρb: the tube fit. A vertex beyond the domain of the branch, where the branch continues as its end section, informs that section: the data around a mouth, where a branch is carried past its own data, keep its extension from standing out of the wall, whatever the extent of its domain. For a branch parametrized by its harmonic angle the wall moves at the vertex's foot pi, its nearest point in the section, along the outward normal ni there: the target is pi+niFrb/wb, at the foot's angle (see below). The step is halved until the misfit does not grow, and the sweeps stop when one decreases the misfit by less than junctions.tolerance. Each wall keeps the knots and the smoothing parameter of a first fit of the branch alone, chosen by generalized cross-validation as for a tube. A sweep evaluates again only the branch it refits: the values fc of the others at every vertex are kept from the previous steps.

Branches that are not star-shaped ​

Every branch is first fitted alone twice, on the vertices it owns: by its polar angle, and by its own harmonic angle. The harmonic angle of a branch is computed as for a tube (Harmonic angle), on the part of the wall where its harmonic field is free: the branch and the branches off it, which are dead ends there, insulated at their outlets. On this sub-surface the field is solved again with its values kept only on the rims where it was prescribed (zero upstream, one beyond the branch), so that a vertex of the labelling that strays across the rim cannot become an extremum inside it, where the seam of the angle, which climbs the field, would stop. The angle is oriented and anchored on the vertices the branch owns.

The branch keeps its harmonic angle when that fit leaves at most junctions.harmonic_residual (one half) of the RMS distance of its data from the polar one: when its sections are far from star-shaped, the polar wall cannot follow them. For a branch whose sections are star-shaped the two fits are about as good, and the polar angle, which keeps the wall a function of the ray, is kept, as in every branch of the circular synthetic presets of Validation and in all but two branches of its aortic arches. Unlike a tube's, the samples of the harmonic fit are not then corrected to their feet on it (Parameter correction): along the outline of a mouth, where the data of a branch end unevenly, the feet on a first fit bunch up, and the refit bulges there.

In the joint fit, the wall of such a branch is moved at the feet of the vertices, and fitted by its distance along the normals there (tangent-distance minimization, Wang, Pottmann & Liu [30]): every sample contributes

(ni⋅(W−qi))2+α(ti⋅(W−qi))2

instead of |W−qi|2, with qi its target and ti the tangent, at the foot's angle. With α=1 this is the ordinary fit, which keeps every foot at its angle; the smoothing penalty resists the uneven parametrization that moving the wall along its normals then makes, and the sweeps converge slowly. With a small α (0.01) the wall slides along itself instead: on the crescent bifurcation the joint fit converges in 10 sweeps rather than 42. The radial and tangential components are then coupled (the normal turns between the frames of φ), and the normal equations are block-banded with blocks of both components.

In a cohort the angles of a branch must be of the same kind in every subject to correspond: a branch is parametrized by its harmonic angle in all of them when at least junctions.harmonic_subjects (half) of the subjects, encoded alone, chose it, and by its polar angle in all of them otherwise. The harmonic angle of a trunk feels the mouths of its side branches, which differ from subject to subject, so it corresponds less well between subjects than the polar angle carried along the centerline, and it is used only where it is needed.

The residual ​

The distance from a vertex to the model is |F|r (with r the radius of the branch that dominates there): the radial distance of the tube where one branch is present.

Triangulating the wall ​

The wall of a branched vessel is the zero set of F. It is triangulated by marching tetrahedra (Doi & Koide [23]) on a grid of the requested spacing: every grid cube is split into six tetrahedra around its main diagonal, and each tetrahedron the surface crosses contributes one or two triangles. Unlike methods that place one vertex per cube, every sign pattern has a single reading, so the result is a manifold. The grid is not swept: the surface is followed from cubes on the wall of every branch to the neighbours across every face it crosses (Wyvill et al. [26]), so F is evaluated only near the wall, and the cost grows with its area rather than with the volume around it. At each point only the branches whose centerline passes within reach are evaluated. The vertices are then moved onto the surface along the gradient of F, and the mesh is cut open at the inlet and the outlets. An unbranched vessel keeps its structured (τ,θ) grid.

The meshes of the shapes of a cohort are instead carried from the mesh of the mean, so that they share its faces (Meshes with the same faces).

Why this construction ​

Branched vessels are usually split into branches at the points where their centerlines part. Antiga & Steinman [16] cut the surface into branches and a bifurcation region with clipping surfaces placed by Voronoi-diagram centerlines [17], and map each branch to longitudinal and circumferential coordinates. Structured meshing methods build the junction from patches. De Santis et al. [19] split a bifurcation into three half-branches, and Decroocq et al. [20] fit a parametric bifurcation template (Zakaria et al. [21]) from centerlines and sweep it into a hexahedral mesh. These need the junction to follow a template, smooth the apex afterwards, or leave the junction outside the model.

furca keeps one model for the whole tree:

  • Every branch is a tube with the same meaning as an unbranched vessel, so the coordinates, the wall maps, the sections and the parameters of the tube carry over to every branch, and an unbranched vessel is the tree of one branch.
  • The junction is part of the fitted surface. It is the smooth union of the branch lumens (an implicit blend in the spirit of Ricci [22]), fitted jointly to the data, and not a separate patch or template. The apex comes out of the data as the saddle of a harmonic field, with no tolerance to choose.
  • The surface is smooth wherever the branches are and has no seams, because F is a smooth function.
  • The tree is found with the same tools as the tube: harmonic fields on the wall and their level rings, which never cross.

Limitations ​

  • Trunk or bifurcation: each junction whose children are close in width costs one more fit of the vessel, and the readings are tried one junction at a time.
  • Branches need sections: a branch, or a stretch between two junctions, needs at least four closed level rings. Junctions closer together than that are not resolved.
  • Sections that are not star-shaped are followed by the harmonic angle of their branch, but the joint fit of such a branch costs more: every evaluation of its fb seeks a nearest point. A branch that is star-shaped in most of a cohort and not in a few of its subjects is parametrized by its polar angle in all of them, and fitted poorly in those few.
  • Sacs that lead nowhere, such as an aneurysm at the apex of a junction, belong to no branch; they are modelled as saccular lobes, found where the flow stagnates.
  • Volume meshes of junctions are not available yet.
  • Cohorts need the same tree in every subject (see Cohorts).

furca — parametric models of blood vessels.