Computation of a probabilistic statistical shape model in a maximum-a-posteriori framework.

Area of Science:

• Computational geometry
• Statistical modeling
• Medical image analysis

Background:

• Establishing shape correspondence is crucial for analyzing shape and variability.
• Manual landmarking is time-consuming and may not capture complex shape variations.
• Existing statistical shape models (SSMs) often rely on predefined correspondences, limiting their ability to handle complex shapes.

Purpose of the Study:

• To develop a novel method for representing mean shapes and variability models using probabilistic correspondence.
• To create a statistical shape model (SSM) that automatically computes correspondences between observations.
• To improve the modeling of shape details and differences in statistical shape analysis.

Main Methods:

• Affine transformation using Expectation-Maximization Iterative-Closest-Points (EM-ICP) registration for initial matching.
• A maximum-a-posteriori (MAP) framework to compute SSM parameters for optimal model adaptation.
• Global optimization of MAP explanation with respect to observation and generative model parameters for efficient, closed-form solutions.

Main Results:

• The probabilistic SSM demonstrated superior performance in modeling shape details compared to traditional SSMs based on one-to-one correspondences and PCA.
• Experiments on synthetic non-convex shapes and real putamen data showed improved generalization ability and specificity.
• The probabilistic SSM achieved approximately 25% smaller Hausdorff distance in generalization ability, indicating better modeling of shape differences.

Conclusions:

• The developed probabilistic SSM is efficient and advantageous for shape analysis.
• This approach excels at capturing fine shape details and variations.
• Probabilistic correspondence offers a more robust method for building statistical shape models.