Skip to content

Cohorts: alignment and shape models ​

Correspondence through a shared structure ​

An encoding is a feature vector

x=(c0,…,cnc−1, u(0), a0,…,anc−1, c00,…,cnτ−1,nθ−1)∈R5nc+3+nτnθ

(centerline control points, angular reference, section tilt coefficients ai=(ai,bi), wall coefficients) together with a structure: the bases with their knots, the wall domain and the frame sampling. Two feature vectors can be compared component by component only if their structures are equal; then component k refers to the same position along the centerline, or the same (τ,φ) neighbourhood of the wall, in both vessels. This is the correspondence that statistical shape analysis needs, and here it comes from the parametrization itself rather than from a separate registration step.

Branched vessels ​

A branched vessel is a tree of tubes, and its feature vector is that of every branch, each followed by its wall domain [τa,τb], and then what shapes its junctions: the radius rb of every branch, the windows of its ends, and the apex of every junction. Two branched vessels correspond when they have the same tree (the same junctions, each read as a trunk or as a bifurcation, with the branches in the same order) and the same structure in every branch. Then a component refers to the same branch and, within it, to the same place relative to its centerline and its junctions. An unbranched vessel is the tree of one branch: its features are those of its tube and its domain, which in a cohort is the same for every subject.

The angular origin of a child is carried over from its parent at the junction by the smallest rotation that takes the parent's tangent to the child's (see Sections of a branch), so θ means the same around a branch in every subject whose junction looks alike.

A shared structure ​

encode_cohort enforces a shared structure:

  1. Tree. Every subject is encoded first as it would be alone. Their trees must be the same; the encoding then keeps the reading of every subject's junctions.
  2. Sizes. Sizes given as rules are resolved for every subject, and the cohort uses the median, branch by branch.
  3. Domain. At the inlet and the outlets, every wall ends at the boundary centroids: τ=0 and τ=1 at the free ends of every subject (end mode centroid). Where a branch meets a junction its domain is a feature.
  4. Knots. With adaptive knots, the normalized variation profiles of all subjects (pilot fits on uniform knots) are averaged and a single knot vector is placed from the mean profile (Adaptive knots), for every branch; then all subjects are refitted on them.

Generalized Procrustes alignment ​

Feature vectors also contain position and orientation, which are usually irrelevant. Let Xi∈Rm×3 be m corresponding surface points of subject i: the wall at a grid of (τ,φ) values over the wall domain of every branch (alignment.tau_samples rings, shared among the branches in proportion to their mean length, with alignment.theta_samples points each). Generalized Procrustes analysis (Gower [11]) finds similarities siRip+ti minimizing

∑i‖siXiRi⊤+1ti⊤−X¯‖F2

by alternating: (i) centre every Xi; (ii) align each to the current mean with the optimal rotation (and scale), obtained from the SVD of X¯⊤Xi with the sign correction of Umeyama [12]; (iii) recompute the mean. With scaling, the mean is renormalized to the average size of the subjects to fix the gauge. Iteration stops when the mean moves less than alignment.tolerance times its RMS radius.

The similarity is then applied exactly to each encoding: control points transform as points, the reference direction rotates, the rotation-minimizing frame follows the centerline (it is equivariant under rotations), the tilt is relative to that frame and does not change, and wall coefficients scale by si. The radii of the branches scale by si and the apexes transform as points; domains and windows, values of τ, do not change.

Principal component analysis ​

With aligned feature vectors x1,…,xN, mean x¯ and centred data matrix Xc, the modes φk and variances λk are the eigenpairs of the sample covariance Xc⊤Xc/(N−1). Since N is much smaller than the feature length, they are computed from the N×N Gram matrix XcXc⊤/(N−1)=UΛU⊤, with φk∝Xc⊤uk (Cootes et al. [13]). The model keeps the leading modes explaining model.variance_fraction of the total variance, and represents shapes as

x(b)=x¯+∑kbkλkφk,

with b in standard deviations. A subject's scores are bk=φk⊤(xi−x¯)/λk.

Decoding x(b) re-normalizes the reference direction (an average of unit vectors is not unit); every other component is used as is.

Explained variance of a synthetic cohort

Explained variance of the synthetic cohort of the user guide: four parameters vary, four modes explain all the variance.

Meshes with the same faces ​

The walls of unbranched shapes are triangulated on their (τ,φ) grid, so their meshes correspond vertex by vertex. A branched wall is triangulated as the zero set of the union of its branches (Triangulating the wall), whose faces depend on the shape. Instead, the triangulation of the mean is carried onto every other shape of the cohort:

  1. Every vertex keeps the branch that dominates it in the mean, the angle φ of the nearest point of that branch's wall in its section, and its signed distance from that point, in radii of the branch (its wall coordinates).
  2. Its τ keeps its place relative to the anchors of the branch: the ends of its windows and of its wall domain, which differ from shape to shape. Between anchors τ is mapped linearly, beyond them it is shifted (along the extensions it is still a length).
  3. Around the mouth of every side branch, where its centerline becomes fully present, the coordinates of the branch it leaves follow the mouth: they are shifted by the displacement of the mouth in that branch's (τ,φ), weighted by a smooth step that falls from one at the mouth to zero three side-branch radii away.
  4. The vertex is placed by the coordinates of its branch in the shape (at that distance along the normal of the section curve), and moved onto the shape's wall along the gradient of the union. Where the junctions of the two shapes differ most (in a crotch between close branches) some faces can fold; their vertices are moved to the mean of their neighbours and back onto the wall.

The result has the faces of the mean's mesh and lies on the shape's wall, except in small patches in the crotches between close branches (Validation). The furca model modes meshes and the shape views of the application are built this way, so that a shape can be animated by moving the vertices of one mesh.

furca — parametric models of blood vessels.