Skip to content

Vessel coordinates ​

Definition ​

The cross-section at τ is a plane through the centerline point c(τ) with unit normal n(τ). A point p has vessel coordinates (τ,θ,ρ) when it lies in the section at τ and

p=c(τ)+ρ(cos⁡θU(τ)+sin⁡θV(τ)),ρ≥0, θ∈[0,2π),

where (U,V) is an orthonormal basis of the section plane. The longitudinal coordinate is therefore a root of the plane offset

g(τ)=n(τ)⋅(p−c(τ)).

In the classical choice the sections are the normal planes of the centerline, n=t; then g=0 is the foot-point condition p−c(τ)⊥t(τ) and (U,V) is the rotation-minimizing frame (u,v). furca uses by default tilted sections that stay valid in tight bends (frame.sections = "harmonic"):

n(τ)=t+au+bv∥t+au+bv∥,

with the tilt (a,b)(τ) a pair of cubic splines on the centerline basis: a and b are the slopes of the section plane in the centerline frame. The in-plane axes are the frame axes carried along by the smallest rotation Rt→n taking t to n,

U=Ru,V=Rv,Rx=(t⋅n)x+(t×n)×x+(t×n)((t×n)⋅x)1+t⋅n,

so that θ=0 keeps its meaning and zero tilt gives back normal sections exactly.

Loading…

The sections of a real descending aorta along its length. Move the slider: the section is shown on the vessel (left) and in its own plane (right) with the axes U, V, the encoded wall W(τ,⋅) (red), the mean radius (dashed) and the input vertices near the section (blue).

When the coordinates are a chart ​

Differentiating p=c+ρe (with e the unit radial direction, e⊥n) shows that the map (τ,θ,ρ)↦p has Jacobian determinant

J=ρ(n⋅c′−ρn′⋅e)=ρ(n⋅c′)(1−χ),χ(τ,θ,ρ)=ρn′⋅en⋅c′,

and that g′(τ)=−J/ρ. The fold χ measures how fast the section planes turn relative to how fast they advance: neighbouring sections meet where χ=1. For normal sections n′=κ∥c′∥N and χ=ρκcos⁡(θ−θN), the familiar condition ρκ<1 of the tubular neighbourhood theorem. Wherever χ<1, g is strictly decreasing, so the section through each point is unique and depends smoothly on it.

χ is linear in the point's position in its plane, so it is defined at any point (x,y) of a section, χ=(xU+yV)⋅n′/(n⋅c′). The encoder reports the largest fold over the encoded wall, χ at W(τ,φ) on a grid of frame.fold_samples points (max_fold). The fold of a point of a section is a convex combination of the folds of points of its wall (it is linear, and every point of a section lies on a segment between two wall points in a line through it), so a value below one on the wall means that no two sections meet anywhere inside the encoded vessel, and the reconstruction cannot fold. The fold at the input vertices would not tell: every vertex is assigned the root of g on the decreasing branch, where χ<1 by construction, even where sections cross.

Wall coordinates ​

The section through a point and its polar coordinates (θ,ρ) in it are defined for any point near the centerline. Relative to the wall, a point of the section at τ with coordinates (x,y) has the wall coordinates (τ,φ,d): the harmonic angle φ of the nearest point of the encoded section to it, and the signed distance d from that point, along the outward normal of the section curve (negative inside). On the wall d=0; near it, (φ,d) is a chart of the section plane up to the distance where the normals of the section curve meet.

Tilted sections ​

Normal sections fail in bends tighter than the vessel: the centerline's normal planes all pass through the centre of curvature, which then lies inside the vessel. The fix must turn the sections more slowly than the tangent on the inner side of the bend, while keeping them close to normal elsewhere. furca obtains them from the same harmonic field u that gives the centerline.

Disjoint planar sections cannot cross inside the vessel. Let two planes cut the wall along disjoint closed curves γ1, γ2 that bound the sections D1, D2. If the planes met inside the vessel, their common line would cross both D1 and D2 along intervals inside the lumen whose endpoints lie on the wall, that is, on γ1 and on γ2. Both intervals are the same maximal interval of the line inside the lumen, so the curves would share a point, a contradiction. A family of disjoint rings, which the level sets of u are by the maximum principle, therefore spans a family of planes that foliates the lumen, however tight the bend.

Level rings are not exactly planar, so furca takes for each ring its vector area

A=12∮x×dx,

whose direction is the normal of a planar ring and, for any closed curve, the normal of the plane onto which it projects with the largest area. The rings are oriented consistently with ∇u. The boundary loops are level sets as well (u=0,1), so oblique end cuts become end sections. Each ring gives a slope (ai,bi)=(A⋅u,A⋅v)/(A⋅t) at its centerline parameter ti. The tilt is the smoothing spline through these slopes with a physical cut-off length ℓt (frame.tilt_stiffness ×r¯, default 0.3 mean radii) and the third-order penalty of the centerline:

mina,b ∑iLN‖(a,b)(ti)−(ai,bi)‖2+ℓt6∫0L‖(a,b)‴(s)‖2ds.

The guarantee that sections do not cross belongs to the planes of the rings themselves; the smoother the fitted tilt, the further it departs from them. In a crease, where neighbouring rings nearly touch, their planes nearly meet at the wall and the fold approaches one: there the tilt must follow the rings closely. The tilt is therefore smoothed on a shorter scale than the centerline. On the tight-bend benchmark, a cut-off of 0.3 radii keeps the fold of the encoded wall below one up to ρκ=1.5 and at every wall resolution; with the cut-off of the centerline the fold exceeds one at the crease of the tightest bend.

On the tight-bend benchmark the planes of the level rings cover every sampled point of the lumen exactly once for ρκ up to 1.5, whereas normal planes cover 2.6–5% of the lumen more than once. The tilt stays below 10°. In straight segments the rings are nearly normal and the tilt vanishes, except within about one radius of an oblique end cut, where it adapts to the cut.

Computation ​

  1. Initial guess. The centerline is sampled at the frame.samples + 1 nodes of the frame table and stored in a k-d tree. The nearest sample gives τ0, which avoids starting on a different part of a curving centerline.
  2. Bracketing. Since g decreases in the valid region, a point with g(τ0)>0 has its section at larger τ. Steps of doubling length from τ0 in that direction bracket a sign change of g.
  3. Refinement. Newton steps τ←τ−g/g′, with g′=n′⋅(p−c)−n⋅c′, are kept inside the bracket and replaced by bisection whenever they would leave it. Convergence is quadratic, to 10−14 in τ.
  4. Angle and radius. The section frame at τ gives θ and ρ.

n′ is evaluated analytically. With m=t+au+bv and the rotation-minimizing property u′=−(u⋅t′)t (likewise for v),

m′=t′+a′u+b′v−(au⋅t′+bv⋅t′)t,n′=m′−(m′⋅n)n∥m∥.

Beyond the ends ​

Beyond its ends the centerline continues along its end tangents,

c(τ)=c(0)+τLt(0)(τ<0),c(τ)=c(1)+(τ−1)Lt(1)(τ>1),

with L the centerline length, and the section plane and its axes are held at their end values. There g is linear in τ, and the same root finder applies unchanged. The extended centerline is C1, so vessel coordinates are continuous across the ends, and every point near the vessel has coordinates, including points beyond an oblique cut.

Reproducibility ​

For a given input, parameters and number of threads, results are bit-for-bit identical: parallel accumulations are done in fixed-size chunks combined in a fixed order. With a different number of threads, the dense linear algebra partitions its work differently and results differ at rounding level (observed below 10−7 relative to the radius).

furca — parametric models of blood vessels.