Globally Optimal Label Fusion with Shape Priors.

Area of Science:

• Neuroimaging
• Computational Anatomy
• Medical Image Analysis

Background:

• Multi-atlas label fusion (MALF) is effective for image segmentation.
• Graph-based methods offer global optimality guarantees.
• Integrating these approaches can enhance segmentation accuracy.

Purpose of the Study:

• To propose GOLF, a novel method combining graph-based segmentation and MALF.
• To incorporate shape priors into the label fusion process.
• To achieve globally optimal solutions for multi-label segmentation tasks.

Main Methods:

• GOLF integrates shape priors with MALF.
• It leverages posterior maps from MALF for improved accuracy.
• The method is demonstrated for joint segmentation of subcortical structures.

Main Results:

• GOLF significantly outperforms FreeSurfer and FIRST across all 8 segmented structures.
• Comparisons with a standard MALF approach highlight GOLF's component contributions.
• The framework provides highly accurate joint segmentation of caudate, putamen, globus pallidus, and nucleus accumbens.

Conclusions:

• GOLF offers a significant advancement in brain image segmentation.
• The proposed method achieves superior accuracy and global optimality.
• It provides a robust framework for joint segmentation of multiple subcortical structures.