Skip to content

Centerline ​

Definition ​

The harmonic field on the wall has level sets that are rings around the vessel and never cross, however tightly the vessel bends. Their length-weighted centroids p1,…,pN (preceded and followed by the centroids of the two boundary loops) sample the axis of the vessel. The centerline is the cubic B-spline curve minimizing

E[c]=∑i=1NLN‖c(ti)−pi‖2+ℓ02k∫0L‖c(k)(s)‖2ds,

where ti are the chord-length parameters of the centroids, L is the length of the vessel, s arc length and k the order of the penalty (centerline.penalty_order, default 3). With the parameter proportional to arc length, ∫∥c(k)(s)∥2ds=L1−2k∫01∥c(k)(t)∥2dt, computed exactly with the matrix ∫Bi(k)Bj(k) of the basis (B-splines). The minimizer is a linear least-squares problem, solved once.

The cut-off length ℓ0. For a deviation of wavelength ℓ and amplitude a, the data term costs about a2 per unit length and the penalty ℓ02ka2/ℓ2k: the two balance at ℓ=ℓ0, whatever k. The centerline therefore follows the axis on scales longer than ℓ0 and ignores shorter ones. furca sets ℓ0 = centerline.stiffness ×r¯, with a default of half a mean radius, so that the weight is a physical length rather than a number estimated from the data.

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

The centerline of a real descending aorta: the centroids of the level sets of the harmonic field (the initial points), the fitted curve and its control points. The layers can be shown and hidden.

Why the third derivative ​

The classical smoothing spline penalizes the bending energy, k=2. It biases the centerline towards zero curvature, in two ways:

  • At the ends. The minimizer satisfies the natural boundary condition c″=0: the curve leaves a bend straight. Near the end of a curved segment it departs from the axis, and the sections built on it look tilted with respect to it.
  • Inside a bend. A circular arc of curvature κ costs ℓ04κ2 per unit length, so the fit trades some distance to the data for less curvature and cuts the bend short, by an amount of order (κℓ0)2ℓ0.

With k=3 the penalty is ℓ06∫∥c‴∥2. Constant curvature costs nothing at the ends (the natural condition becomes c‴=0), and a circular arc, for which ∥c‴∥=κ2, costs ℓ06κ4: the bias inside a bend falls by a further factor of order (κℓ0)2. The filtering of short wavelengths is, if anything, sharper.

On synthetic tubes, whose axis is known exactly, the bending penalty left the centerline up to a fifth of a radius away from the axis near the ends of an arch, with spurious section tilts of up to 16°. The third-order penalty keeps it within a few thousandths of a radius (Validation), and centres the sections of the real segments two to three times better.

The number of control points (centerline.size) is chosen generously: the cut-off, not the resolution, decides the smoothness.

Why not generalized cross-validation ​

The first versions of furca chose the smoothing of the centerline by generalized cross-validation, as for the wall. GCV assumes independent errors, but consecutive level-set centroids are strongly correlated (neighbouring rings share most of their geometry). GCV then under-smooths: with few control points this goes unnoticed, but with the resolution needed in tight bends the centerline wiggles, and the error is amplified on the inner side of bends (see below). A physical cut-off has no such failure mode.

Why not iterative centering ​

The first versions also centered the centerline iteratively, moving it until the cross-sections normal to it had no first angular Fourier mode. In tight bends this is ill-posed: the normal sections themselves depend on the curve being moved, sections near the inner side compress (the area element of normal-section coordinates carries a factor 1−ρκcos⁡θ) and eventually cross. The harmonic rings provide sections that do not depend on the curve and never cross, so fitting their centres is both simpler and more robust.

Tight bends ​

The centerline does not have to be less curved than the axis for the coordinates to survive tight bends: the tilted harmonic sections never cross, whatever the curvature of the centerline. It can therefore follow the axis closely. On a benchmark of tubes with a 120° bend, the earlier GCV-smoothed and iteratively centred centerline failed from ρκ≈0.85 on; the harmonic-section centerline with tilted sections stays accurate up to ρκ=1.5; see Validation.

Centring diagnostic ​

After the fit, furca measures how well the centerline is centred in its cross-sections: in bins along the vessel (centerline.offset_bins) whose angular sectors (centerline.offset_sectors) all contain data, it fits ρ≈a0+a1cos⁡θ+b1sin⁡θ by area-weighted least squares; the offset of the section's Fourier centre is (a1,b1). The diagnostics report the mean and largest offset relative to the mean radius (centerline_offset). In straight and gently curved vessels it is a few percent of the radius; in tight bends it grows, because the centroid of a harmonic ring and the Fourier centre of a normal section are different notions of centre once sections are no longer parallel.

Ends ​

Vessel segments are usually cut by planes that are not exactly normal to the centerline. With an oblique cut, part of the wall lies beyond the normal plane at the end of the centerline. Rather than moving the centerline to cover it, which couples the shape of the curve to the cut, furca extends the coordinate system: beyond its ends the centerline continues along its end tangents with a constant frame (see Vessel coordinates). Every point near the vessel then has coordinates, with τ<0 or τ>1 beyond the ends.

The wall is fitted over a longitudinal domain [τa,τb] chosen by ends.mode:

  • boundary (default): the smallest interval containing the τ of every vertex. The end sections are then the normal planes that just contain the whole surface and every vertex is fitted. Near an oblique end the sections are only partly covered by data; the penalty extrapolates the wall smoothly over the uncovered part.
  • centroid: [0,1], so the model ends at the boundary centroids. Vertices with τ outside [0,1] by more than ends.axial_tolerance ×r¯ (in arc length) are not fitted.

On the real test segments, boundary mode reduces the largest distance from the input to the model from 6–23% of the radius to below 2%. The extensions it needs are small, typically 0.05–0.4 radii.

furca — parametric models of blood vessels.