A method and software for segmentation of anatomic object ensembles by deformable m-reps

Area of Science:

• Medical Image Analysis
• Computer Vision
• Computational Anatomy

Background:

• Deformable shape models (DSMs) are a promising approach for automatic image segmentation.
• Successful segmentation requires knowledge of target object geometry, contextual objects, and image intensities relative to geometry.
• Efficient segmentation demands multi-scale deformations, including bending and twisting.

Purpose of the Study:

• To describe a novel deformable shape model called m-reps.
• To present a segmentation method using multi-scale posterior optimization of m-reps.
• To introduce software implementing this segmentation approach for 3D medical images.

Main Methods:

• Utilized explicit geometric models within a Bayesian statistical framework for a priori information.
• Employed posterior optimization to match the DSM to target image data through transformations.
• Developed m-reps, a DSM form allowing multi-scale deformations (translation, bending, twisting, magnification).

Main Results:

• Implemented segmentation software using deformable m-reps, achieving 3D segmentations in minutes.
• Developed software for building and training m-rep models.
• Demonstrated successful segmentation by optimizing m-reps from large to small scales.

Conclusions:

• M-reps provide a robust framework for deformable shape model-based segmentation.
• The described method enables efficient and accurate 3D medical image segmentation.
• The developed software facilitates rapid and effective anatomical structure segmentation.