Multi-Atlas Segmentation with Joint Label Fusion.

Area of Science:

• Biomedical image analysis
• Medical image segmentation
• Computational anatomy

Background:

• Multi-atlas segmentation relies on registering and fusing expert-segmented images (atlases) to label structures in target images.
• Current weighted voting methods use atlas-target similarity but ignore potential shared errors between atlases.

Purpose of the Study:

• To develop a novel label fusion strategy for multi-atlas segmentation that accounts for pairwise dependencies between atlases.
• To improve the accuracy of automated segmentation in biomedical images by minimizing the total expected labeling error.

Main Methods:

• Formulated weighted voting as an error minimization problem, explicitly modeling joint error probabilities between pairs of atlases.
• Approximated pairwise atlas error probabilities using intensity similarity between atlases and the target image in local neighborhoods.

Main Results:

• Demonstrated consistent and significant improvements in segmentation accuracy compared to independent atlas weighting methods.
• Validated the method on hippocampus and hippocampus subfield segmentation tasks in magnetic resonance (MR) images.

Conclusions:

• The proposed method effectively models inter-atlas dependencies, leading to superior segmentation performance.
• This approach offers a more robust solution for multi-atlas segmentation, particularly in complex anatomical regions.