Fusion set selection with surrogate metric in multi-atlas based image segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Computational anatomy

Background:

• Multi-atlas based image segmentation faces challenges with big data, including data heterogeneity and high computational costs.
• Selecting relevant atlases before label fusion is critical for performance and efficiency.

Purpose of the Study:

• To investigate image similarity metrics ('surrogates') as alternatives to geometric agreement metrics ('oracles') for assessing atlas relevance.
• To develop a model for relating surrogates and oracle metrics to guide atlas selection.
• To provide insights into optimal fusion set sizes for effective atlas-based segmentation.

Main Methods:

• An inference model was proposed to link surrogate and oracle geometric agreement metrics.
• The behavior of surrogates in mimicking oracle metrics for atlas relevance ordering was quantified.
• Probabilistic analysis was used to determine optimal fusion set sizes.

Main Results:

• The proposed model effectively quantifies how well surrogate metrics predict oracle metrics for atlas relevance.
• Analytical insights were derived for selecting the optimal number of atlases to include in segmentation.
• The methods were validated using prostate and corpus callosum segmentation datasets.

Conclusions:

• The developed approach enhances the selection of relevant atlases in multi-atlas segmentation, mitigating performance loss and computational overhead.
• This work provides a framework for optimizing atlas selection strategies in big data medical imaging.
• The findings contribute to more accurate and efficient image segmentation through informed atlas selection.