A framework for the merging of pre-existing and correspondenceless 3D statistical shape models

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Statistical Shape Models (SSMs) are crucial for representing anatomical variability in medical imaging.
• Limited training data and the complexity of anatomical structures pose challenges in constructing rich SSMs.
• Existing methods often rely on data decomposition or synthetic models to address data scarcity.

Purpose of the Study:

• To present a flexible framework for merging multiple 3D statistical shape models into a single, integrated model.
• To enhance the representation of anatomically meaningful shape variability without requiring new real datasets.
• To enable the re-use and complementation of pre-existing SSMs when data sharing is restricted.

Main Methods:

• A two-stage approach involving statistical model normalization and integration.
• Surface-based registration is used for normalization to establish point correspondences across eigenspaces.
• A model fusion algorithm is applied to combine normalized models into a unified SSM with improved generalization.

Main Results:

• The framework successfully merged statistical models of cardiac ventricles, L1 vertebra, and caudate nucleus.
• Integration of models from different research centers, imaging modalities (CT, MRI), and correspondences proved effective.
• The resulting unified SSM demonstrated statistically and anatomically meaningful improvements in shape variability representation.

Conclusions:

• The proposed method offers a flexible and effective way to integrate multiple statistical shape models.
• This approach enhances the generalization ability of SSMs by encoding additional anatomical variability.
• The framework holds significant potential for merging pre-existing, multi-modality SSMs, advancing medical image analysis.