Gaussian Process Morphable Models | JoVE Visualize

Area of Science:

Computer Vision
Medical Image Analysis
Statistical Shape Modeling

Background:

Point Distribution Models (PDMs) are widely used for automated image analysis, representing shape variations via principal component analysis (PCA).
PDMs, however, are limited to the linear span of example data, restricting their flexibility in capturing complex shape variations.

Purpose of the Study:

To introduce Gaussian Process Morphable Models (GPMMs) as a generalization of PDMs, offering enhanced flexibility and continuous shape variation modeling.
To demonstrate the applicability of GPMMs in non-rigid registration and model-based segmentation, particularly in medical image analysis and computer vision.

Main Methods:

GPMMs model shape variations using Gaussian processes, represented by Karhunen-Loève expansion components computed via the Nyström method.
A novel, simple algorithm for fitting GPMMs to surfaces or images is presented, enabling non-rigid registration with GPMM-defined regularization.
The approach separates modeling from fitting, allowing diverse registration schemes (multi-scale, hybrid) without altering the fitting algorithm.

Main Results:

GPMMs provide a continuous analog to PDMs, allowing for arbitrary Gaussian process-defined shape variations, including spline-based models without example data.
The proposed fitting algorithm successfully performs non-rigid registration and model-based segmentation, as demonstrated in 3D forearm image segmentation and facial statistical modeling.
The flexibility of GPMMs allows for combining different models, such as extending PDMs with spline models for enhanced shape representation.

Conclusions:

GPMMs offer a powerful and versatile generalization of PDMs, overcoming limitations of linear shape variation modeling.
The proposed method enables flexible, continuous shape modeling and robust non-rigid registration with applications in medical imaging and computer vision.
Open-source availability of the methods facilitates further research and application in automated image analysis and shape modeling.