Deformable image registration by combining uncertainty estimates from supervoxel belief propagation

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Discrete optimization offers advantages for deformable medical image registration over continuous methods.
• High-dimensional medical data complexity makes direct voxel-wise deformation estimation impractical.
• Previous graphical model approaches relied on parameterized models and coarse-to-fine schemes.

Purpose of the Study:

• To develop an accurate, voxel-wise deformable registration method for high-resolution 3D images.
• To eliminate the need for intermediate image warping or multi-resolution strategies.
• To leverage supervoxel layers and belief propagation for robust deformation inference.

Main Methods:

• Representing the image domain using multiple supervoxel layers.
• Inferring deformation regularity using belief propagation and utilizing full marginal distributions.
• Employing minimum spanning trees to model pairwise deformation dependencies.
• Calculating optimal displacements by finding the mode of probability distributions across overlapping supervoxel graphs.

Main Results:

• Demonstrated applicability in challenging intra-patient lung CT motion estimation and MRI brain atlas-based segmentation propagation.
• Achieved continuous-valued sub-voxel motion vectors for lung registration using mean-shift.
• Improved registration accuracy by incorporating displacement uncertainty estimates.
• Enabled fusion of complementary proposals for probabilistic one-to-one image registration.

Conclusions:

• The proposed multi-layer supervoxel approach enables accurate, efficient voxel-wise deformable registration without coarse-to-fine schemes.
• The method effectively handles high-resolution 3D data and improves accuracy through uncertainty estimation.
• This probabilistic framework extends fusion concepts to one-to-one image registration, offering significant advancements.