Discriminative confidence estimation for probabilistic multi-atlas label fusion

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Anatomy

Background:

• Accurate segmentation of anatomical brain structures is crucial for quantitative neuroimaging.
• Manual segmentation is time-consuming and not scalable, necessitating automated methods.
• Multi-atlas segmentation offers state-of-the-art performance but relies heavily on accurate label fusion.

Purpose of the Study:

• To develop a novel probabilistic label fusion framework for multi-atlas brain segmentation.
• To improve the accuracy and robustness of automatic brain structure segmentation.
• To enhance the estimation of atlas contributions to the final segmentation.

Main Methods:

• Proposed a probabilistic label fusion framework using voxel-wise atlas label confidences.
• Estimated maximum likelihood atlas confidences via a supervised approach, modeling local image appearance and segmentation errors.
• Introduced novel label-dependent appearance features for improved confidence estimation and evaluated spatial pooling strategies.

Main Results:

• The proposed framework achieved superior performance compared to state-of-the-art methods for segmenting multiple subcortical brain structures and hippocampi.
• Demonstrated increased robustness to image registration errors.
• The novel label-dependent features significantly improved label fusion accuracy.

Conclusions:

• The developed probabilistic label fusion framework offers a significant advancement in multi-atlas brain segmentation.
• The method provides accurate and robust segmentation, outperforming existing approaches.
• This work contributes to more scalable and reliable quantitative neuroimaging analyses.