Learning image based surrogate relevance criterion for atlas selection in segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Multi-atlas based image segmentation relies on selecting geometrically relevant atlases from large datasets.
• Current image similarity criteria lack clear understanding and optimization for geometric relevance.
• The Big Data era presents challenges with heterogeneous data quality in atlas selection.

Purpose of the Study:

• To develop an image-based surrogate criterion that accurately reflects underlying geometric relevance for multi-atlas segmentation.
• To establish an optimization framework for learning this surrogate relevance criterion.
• To improve the selection of geometrically relevant atlases in image segmentation tasks.

Main Methods:

• Formulated the surrogate relevance criterion learning as an optimization problem.
• Trained the surrogate to be small for similar geometries and large for dissimilar geometries.
• Validated the learned surrogate on corpus callosum segmentation tasks.

Main Results:

• The developed image-based surrogate criterion effectively captures geometric relevance.
• Learned surrogate demonstrated improved performance compared to existing benchmark candidates.
• Experiments confirmed the utility of the surrogate in enhancing segmentation quality.

Conclusions:

• The proposed surrogate relevance criterion offers a significant advancement for multi-atlas image segmentation.
• This method addresses the challenge of selecting geometrically relevant atlases in large, heterogeneous datasets.
• The optimization framework provides a robust approach to learning effective relevance measures for image segmentation.