Landmark and Intensity Based Registration with Large Deformations via Quasi-conformal Maps

1. Introduction.

• Landmark registration matches important features and can handle larger deformations, while intensity registration uses image or surface information without requiring feature landmarks but usually struggles with large geometric deformations.The paper focuses on landmark-based and hybrid registration.Source: §1 · p. 1
• Large deformations can break bijectivity and create overlaps, making registration inaccurate.The introduction identifies diffeomorphic registration under large deformations as a needed capability.Source: §1 · pp. 1–2
• The method minimizes an energy involving a Beltrami coefficient to control registration smoothness and bijectivity, including under very large deformations.The paper reports diffeomorphic image or surface registration with landmark constraints, including large numbers of landmarks.Source: §1 · p. 2
• The proposed landmark-based model is intended to handle very large deformations and large numbers of landmark constraints.The paper also extends this model to hybrid registration combining landmark and intensity information for more accurate registration.Source: §1 · p. 2

2. Previous works.

• Intensity-based methods match image intensities or surface geometry; cited examples include diffeomorphic demons and curvature-based surface registration.The cited survey is, and examples include Vercauteren et al. and Lyttelton et al..Source: §2 · p. 2
• Conformal registration preserves local geometry, while quasi-conformal surface methods permit bounded conformality distortion and can match geometric quantities such as curvature.The passage cites conformal registration work and quasi-conformal work [21].Source: §2 · p. 3
• Earlier landmark methods include thin-plate splines, harmonic maps, and vector-field approaches, with trade-offs in exact matching, bijectivity, map representation, memory, or computational cost.The passage notes that some harmonic methods cannot guarantee bijectivity under many constraints, while vector-field methods can guarantee exact matching but may not represent all surface maps.Source: §2 · p. 3
• Prior hybrid methods combine landmark and intensity information, and the paper proposes a hybrid quasi-conformal method for very large deformations.The cited examples include Christensen et al., Paquin et al., and Chanwimaluang et al.; the passage says existing hybrid techniques work when deformations are not too large.Source: §2 · pp. 3–4

3. Mathematical background.

• A quasi-conformal map is an orientation-preserving homeomorphism with bounded conformality distortion, characterized through a Beltrami equation.For maps from the complex plane to itself, the coefficient satisfies ||µ||∞ < 1.Source: §3 · p. 4
• The Beltrami coefficient measures local deviation from conformality; its value determines the directions and factors of maximal stretching and shrinking.The maximal magnification direction is arg(µ(p))/2 with factor 1 + |µ(p)|, while the orthogonal direction shrinks by factor 1 − |µ(p)|.Source: §3 · pp. 4–5
• Figure 3.2 depicts computing a diffeomorphism’s Beltrami coefficient, reconstructing a quasi-conformal map from it, and displaying the coefficient norm.The figure presents the original mapping, the reconstructed mapping, and the coefficient norm in panels A, B, and C.Source: Figure 3.2 · p. 5

4. Proposed algorithm.

• The Beltrami-coefficient energy controls registration smoothness and bijectivity, enabling quasi-conformal maps for large-deformation image or surface registration.The proposed model seeks a map by minimizing an energy involving a term that measures quasi-conformal distortion.Source: §4 · p. 5

4.1. Proposed model.

• The model minimizes Beltrami-coefficient energy subject to landmark matching and a coefficient bound that guarantees a diffeomorphic map.The energy's first term ensures smoothness, while its second minimizes conformality distortion.Source: §4.1 · p. 6
• The landmark-matching energy has a minimizer in the admissible set, and the combined landmark-and-intensity energy also has a minimizer when both intensity functions are continuous.The stated existence results rely on compactness of the admissible set and, for the combined model, continuity of the intensities.Source: §4.1 · pp. 6–7

4.2.1. Landmark based registration model.

• The landmark-based model alternates between constructing a landmark-constrained map with the Linear Beltrami Solver and updating its Beltrami coefficient to reduce energy.The discrete subproblems use sparse linear systems; the coefficient update is adjusted so landmark mismatch decreases while smoothness is retained.Source: §4.2.1 · pp. 8–9

4.2.2. Hybrid registration model.

• The hybrid model adds image-intensity or surface-curvature matching to landmark constraints and Beltrami regularization to improve registration accuracy.It alternates coefficient and map updates, using the Linear Beltrami Solver to enforce landmark matching.Source: §4.2.2 · pp. 9–11

5.1. Numerical details of LBS.

• The Linear Beltrami Solver computes a quasi-conformal map by discretizing Equation 4.12 into two linear systems for its x- and y-coordinate functions.The map is piecewise linear on each triangular face, where its coordinate gradients determine the Beltrami coefficient.Source: §5 · p. 11, §5.1 · pp. 12, 14
• Boundary conditions may be imposed on the whole boundary or left free, while interior landmark correspondences are enforced as f(pi) = qi.For rectangular domains, the free-boundary example fixes corner values and boundary arcs.Source: §5.1 · p. 14

5.2. Discretization of E-L equations.

• The Euler–Lagrange discretization averages face-based Beltrami coefficients over each vertex’s 1-ring neighborhood and uses a mesh Laplacian.The vertex-defined solution is approximated on each face before the next step.Source: §5.2 · pp. 14–15

5.3. Intensity matching.

• Intensity matching uses the Demons force to find a deformation direction and stabilize the descent, avoiding direct second-order derivatives of µ.The method combines that direction with a descent direction for the term R S1 |µ − νn|2.Source: §5.3 · p. 15

5.4. Multiresolution scheme.

• The multiresolution scheme registers coarsened images first, then interpolates the map to finer scales as the initialization for each subsequent registration.The resulting map matches intensity and satisfies landmark constraints at the original resolution.Source: §5.4 · p. 16

6.1. Landmark based registration.

• With 78 landmark pairs and in a separate example with very large prescribed displacements, the registrations remained bijective without overlaps or flips.The examples use synthetic rectangles; Fig. 6.2 depicts the large-deformation landmark match.Source: §6.1 · p. 16, Figure 6.2 · p. 17
• For brain cortical surfaces, six sulcal landmarks were matched where conformal registration failed; registered surfaces also formed a mean with sulcal features preserved.The mean was computed from 10 cortical surfaces registered using six sulcal landmarks.Source: §6.1 · p. 17, Figure 6.3 · p. 18, Figure 6.5 · p. 19
• Across tiny, moderate, and large deformations, the proposed method produced non-overlapping landmark matches in the least time; harmonic maps and TPS overlapped, while LDDMM was more expensive.The comparison is reported across the three deformation sizes; the table caption identifies the comparison scope.Source: §6.1 · p. 17, Table 6.1 · p. 17

6.2. Hybrid registration.

• In the three-layer synthetic hybrid example, optimization reduced energy at the coarsest layer and reached an optimal registration at the finest layer after about 10 iterations.The deformed image closely resembled the target, with landmarks consistently matched and the registration bijective.Source: §6.2 · p. 19, Figure 6.6 · p. 19, Figure 6.7 · p. 20
• For the larger-deformation synthetic I-to-C image pair, the hybrid registration closely resembled the target while consistently matching landmarks and remaining bijective.Source: §6.2 · p. 20, Figure 6.9 · p. 21
• <1% intensity mismatching error was reported for the human-hand hybrid registration, whose deformed image closely resembled the target unlike either single-source approach.The comparisons show landmark-only and intensity-only results as well as the hybrid result.Source: §6.2 · pp. 20–21, Figure 6.11 · p. 22
• <1% intensity mismatching error was also reported for brain MRI hybrid registration; the result resembled the target, while intensity-only registration left the tumor unmoved.Source: §6.2 · p. 21, Figure 6.13 · p. 23
• On teeth surfaces, landmark-only registration mismatched curvature while the hybrid result matched it; on face surfaces, corresponding regions matched and curvature mismatch was iteratively minimized.Source: §6.2 · pp. 22–23

• The method minimizes an energy involving a Beltrami coefficient that measures conformality distortion and controls registration bijectivity and smoothness.Source: §7 · p. 23
• In the two-tooth example, landmarks are labeled, but mapped mean curvature is not consistently matched in the boxed regions.Source: Figure 6.15 · p. 24, Figure 6.14 · p. 24
• Experiments on synthetic and real data report diffeomorphic registrations with least local geometric distortion, guaranteed bijective under large deformations or many landmark constraints.Source: §7 · p. 24
• The model supports landmark-based and hybrid registration; future work targets high-genus surfaces and more medical-imaging applications for disease analysis.Source: §7 · p. 24