Brain Extraction Using Label Propagation and Group Agreement: Pincram

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate brain delineation on magnetic resonance (MR) images is crucial for neuroimaging analysis.
• Existing brain segmentation methods often suffer from laboriousness, inconsistency, and dependence on specific training data.
• There is a need for automated, accurate, and versatile brain segmentation tools.

Purpose of the Study:

• To present Pincram, an automatic and versatile method for precise adult brain labeling on T1-weighted 3D MR head images.
• To overcome the limitations of current manual and semi-automatic segmentation techniques.

Main Methods:

• Pincram employs an iterative refinement approach using image registration to propagate labels from multiple atlases to target MR images.
• A consensus label is generated at each refinement level, constraining subsequent boundary searches to the neighborhood of the consensus boundary.
• A novel self-monitoring feature generates a 'success index' to indicate the accuracy of the segmentation output.

Main Results:

• Pincram achieves high accuracy, with a Jaccard coefficient greater than 0.95 and a Dice similarity coefficient greater than 0.97 on typical data.
• Independent evaluation on the Segmentation Validation Engine demonstrates superior performance compared to many state-of-the-art methods.
• The 'success index' provides a reliable measure of segmentation accuracy.

Conclusions:

• Pincram offers an accurate, automated, and versatile solution for brain segmentation in T1-weighted MR images.
• The method's iterative refinement and self-monitoring features enhance reliability and usability.
• Pincram is available as open-source software, facilitating wider adoption in neuroimaging research.