Gaussian Process Morphable Models

• GPMMs are statistical models representing shape variations as continuous Gaussian processes that unify PCA-based SSMs and spline methods.
• They employ kernel-driven methods to model non-rigid deformations with high flexibility and precise spatial regularization.
• Applications include robust medical image registration, segmentation, and augmented morphable model construction with built-in uncertainty quantification.

Gaussian Process Morphable Models (GPMMs) are a class of statistical models that represent shape variation as realizations of a continuous Gaussian process over geometric domains, unifying traditional principal component analysis (PCA)-based statistical shape models (SSMs), spline-based models, and other priors in a fully probabilistic, kernel-driven framework. GPMMs furnish a continuous-domain, highly flexible family of priors and generative models for non-rigid deformations, with broad applications in medical image registration, morphable model construction, and statistical shape analysis (Lüthi et al., 2016).

1. Theoretical Foundation and Model Structure

Let Ω ⊂ ℝ³ denote the geometric domain of interest (e.g., the support of a reference surface or a volumetric region). The unknown deformation field $u : Ω \to \mathbb{R}^3$, which maps points from a reference template to new shapes, is modeled as a realization of a matrix-valued Gaussian process:

$u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$

Here, $\mu: Ω \to ℝ³$ is the (optional) mean deformation function (often set to zero for registration), and $k: Ω \times Ω \to ℝ^{3 \times 3}$ is a positive-definite covariance (kernel) function encoding spatial dependencies, smoothness, and scale.

The Karhunen–Loève expansion provides the principal representational mechanism:

$$u(x) = \mu(x) + \sum_{i=1}^∞ α_i \sqrt{λ_i} φ_i(x) \qquad α_i \sim \mathcal{N}(0, 1)$$

where $(λ_i, φ_i)$ are the eigenpairs of the integral operator associated with $k$:

$\int_{Ω} k(x, x') φ_i(x) dρ(x) = λ_i φ_i(x')$

In practical computation, the expansion is truncated to $r$ terms, yielding a finite-dimensional parametric model:

$û(x) = \mu(x) + \sum_{i=1}^r α_i \sqrt{λ_i} φ_i(x)$

The approximation error corresponds to the sum of the discarded eigenvalues $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$0 (Lüthi et al., 2016).

2. Kernel Design and Composition

The choice and construction of the kernel $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$1 determine the specificity, expressiveness, and regularization properties of the GPMM. GPMMs generalize SSMs by allowing arbitrary kernels, enabling the synthesis of models that:

• No training data (pure priors): Use analytical kernels such as Gaussian (squared-exponential), Matérn, or B-spline kernels to impose generic smoothness or multi-scale regularity.
• Empirical (data-driven) models: Define $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$2 as the empirical covariance function from training deformations:

$u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$3

• Hybrids: Compose kernels algebraically via summation, multiplication, localization, or spatial mixtures, producing multi-scale, spatially-varying, or region-specific priors.

For example:

• Multi-scale mixture: $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$4
• Bias-corrected model: $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$5
• Spatially varying mixture: Partition $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$6 into subregions $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$7 with weights $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$8, using $u(x) \sim \mathrm{GP}(\mu(x),\, k(x, x'))$9

All these constructions leverage the closure of positive-definite kernels under addition and multiplication, allowing flexible model composition (Lüthi et al., 2016, Gerig et al., 2017).

3. Numerical Approximation: Nyström Method and Eigenfunction Computation

Closed-form eigenpairs for the integral operator of $\mu: Ω \to ℝ³$0 are only available for special choices of $\mu: Ω \to ℝ³$1. In general, the leading $\mu: Ω \to ℝ³$2 eigenpairs are approximated numerically via the Nyström method:

1. Sample $\mu: Ω \to ℝ³$3 points $\mu: Ω \to ℝ³$4 from $\mu: Ω \to ℝ³$5.
2. Build the block kernel matrix $\mu: Ω \to ℝ³$6: each $\mu: Ω \to ℝ³$7.
3. Compute the top $\mu: Ω \to ℝ³$8 eigenpairs $\mu: Ω \to ℝ³$9 of $k: Ω \times Ω \to ℝ^{3 \times 3}$0 (e.g., via randomized SVD).
4. Extend to the continuous domain using

$k: Ω \times Ω \to ℝ^{3 \times 3}$1

and normalize; set $k: Ω \times Ω \to ℝ^{3 \times 3}$2.

This process decouples the continuous modeling from the numerical eigenpair extraction, allowing arbitrary resolution at the fitting stage and modular model design (Lüthi et al., 2016).

4. Model Fitting and Registration Procedures

GPMMs support unified and modular fitting algorithms for surface or image data. The parameter vector $k: Ω \times Ω \to ℝ^{3 \times 3}$3 (the coefficients in the Karhunen–Loève basis) is optimized via Maximum a Posteriori estimation:

$k: Ω \times Ω \to ℝ^{3 \times 3}$4

where:

• $k: Ω \times Ω \to ℝ^{3 \times 3}$5 with $k: Ω \times Ω \to ℝ^{3 \times 3}$6
• $k: Ω \times Ω \to ℝ^{3 \times 3}$7 is the data fidelity term, e.g.,
  • For surfaces: $k: Ω \times Ω \to ℝ^{3 \times 3}$8 using closest-point residuals
  • For images: $k: Ω \times Ω \to ℝ^{3 \times 3}$9

Gradient-based methods are typically employed for the optimization. Landmark correspondences or region-specific constraints can be accommodated by conditioning the GP prior via standard Gaussian process regression, yielding closed-form expressions for posterior means and covariances (cf. Rasmussen & Williams, 2006) (Lüthi et al., 2016).

5. Applications and Clinical Examples

GPMMs have demonstrated effectiveness across medical imaging, computational morphometrics, and graphics. Notable applications include:

• Model-based segmentation: In 3D forearm CT, GPMMs with anisotropic multi-scale kernels attained improved compactness and generalization compared to standard SSMs, using fewer modes to cover equivalent variance and yielding superior leave-one-out surface fitting metrics.
• Non-rigid registration: Utilizing multi-scale, spatially-varying, or hybrid GPMMs provides more accurate and parameter-efficient fits compared to classical single- or multi-resolution B-spline algorithms (e.g., 500 vs. ~38,000 parameters), especially in bone surface registration tasks.
• Augmented Active Shape Models (ASM): Extending traditional ASM with smooth or localized prior kernels, and posterior GP conditioning on sparse landmarks, achieves measurable improvements in registration accuracy (Lüthi et al., 2016).

6. Generalization, Related Models, and Extensions

GPMMs subsume and unify a range of deformation priors:

• PCA-based SSMs: By setting $$u(x) = \mu(x) + \sum_{i=1}^∞ α_i \sqrt{λ_i} φ_i(x) \qquad α_i \sim \mathcal{N}(0, 1)$$0 as the empirical covariance of training deformations, GPMMs reduce to classical SSMs, but their continuous formulation avoids fixed landmark discretization and admits arbitrary sampling.
• Spline models: Using analytical kernels (e.g., B-splines, Gaussian) with appropriate hyperparameters replicates traditional smoothness regularization, enabling pure smooth priors without any training data.
• Hybrid and compositional models: Algebraic combination of empirical and smoothness kernels integrates learned and prior-informed characteristics, supporting bias correction and coverage beyond the linear training span.
• Custom kernels: Application-specific priors (e.g., mirror symmetry for faces, anisotropic kernels for long bones) can be encoded directly in $$u(x) = \mu(x) + \sum_{i=1}^∞ α_i \sqrt{λ_i} φ_i(x) \qquad α_i \sim \mathcal{N}(0, 1)$$1 (Lüthi et al., 2016, Gerig et al., 2017, Ploumpis et al., 2019).

7. Advantages and Practical Implications

Key advantages of GPMMs include:

• Modeling flexibility: The choice and composition of kernels provide direct control over spatial regularization, multi-scale effects, region specificity, and incorporation of anatomical priors, facilitating highly expressive models.
• Separation of modeling and inference: The GPMM framework decouples prior modeling, kernel design, and basis extraction from downstream fitting and optimization, simplifying experimentation with new priors or constraints.
• Probabilistic interpretation: Modeling deformations as GP realizations enables built-in uncertainty quantification and supports extensions to regression, marginalization, and Bayesian inference frameworks (as in DMFC-GPMs and mesh morphing GPs) (Lüthi et al., 2016, Fouefack et al., 2021, Casenave et al., 2023).
• Computational scalability: Karhunen–Loève truncation and Nyström approximations keep computational demands tractable even for large domains and allow adaptation to new target resolutions at inference time (Lüthi et al., 2016, Casenave et al., 2023).

Experiments confirm that GPMMs equipped with application-specific kernels outperform both generic smoothness priors and purely data-driven SSMs in challenging tasks such as high-fidelity registration, segmentation, and full-head morphable modeling (Lüthi et al., 2016, Ploumpis et al., 2019). The modularity, generalization ability, and compositional flexibility of GPMMs have established them as a foundational paradigm in contemporary statistical shape modeling and non-rigid registration.