Spine segmentation using articulated shape models.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Prior shape information enhances medical image segmentation.
• Existing methods often focus on individual object segmentation, neglecting inter-object relationships.
• Simultaneous segmentation of object constellations can benefit from joint shape and spatial relation priors.

Purpose of the Study:

• To present a two-scale framework for modeling and segmenting object constellations, using the spine as a representative example.
• To leverage both individual object shape priors and spatial relationships for improved segmentation.
• To enhance the accuracy of medical image segmentation for complex anatomical structures.

Main Methods:

• A two-scale framework integrating global spine shape (consecution of vertebra coordinate systems) and local vertebra models (triangulated surface meshes).
• Model adaptation through attraction to image features, constrained by learned shape priors.
• Simultaneous segmentation of the entire object ensemble, considering spatial interdependencies.

Main Results:

• Achieved an average segmentation accuracy of 1.0 mm on ten thoracic CT images.
• Demonstrated improved segmentation results compared to previous approaches for object constellations.
• Successfully segmented the spine as an example of a complex anatomical object constellation.

Conclusions:

• The proposed two-scale framework effectively models and segments object constellations like the spine.
• Integrating global and local shape information with spatial relations significantly improves segmentation accuracy.
• This approach offers a robust solution for enhancing medical image segmentation of complex structures.