A Bayesian model of shape and appearance for subcortical brain segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Anatomy

Background:

• Accurate segmentation of subcortical structures in human brain MRI is challenging due to poor and variable contrast.
• Existing methods often require additional information or complex fitting procedures for successful segmentation.

Purpose of the Study:

• To develop and assess a novel automated method for segmenting 15 subcortical brain structures using a Bayesian framework.
• To improve segmentation accuracy and provide a direct measure of geometric change for clinical applications.

Main Methods:

• Utilized Active Shape and Appearance Models within a Bayesian framework, integrating manually labelled anatomical data.
• Trained the model on 336 manually-labelled T1-weighted MR images.
• Assessed the method quantitatively using Leave-One-Out testing and qualitatively on an independent clinical dataset.

Main Results:

• Achieved median Dice overlaps between 0.7 and 0.9, demonstrating high segmentation accuracy.
• The Bayesian approach efficiently calculated conditional probabilities, avoiding common technical issues.
• The method provides a local measure of geometric change, independent of tissue classification or smoothing.

Conclusions:

• The proposed automated segmentation method offers a robust and accurate solution for subcortical structure segmentation in brain MR images.
• This approach is comparable or superior to existing automated methods and has been implemented as FIRST in the FSL package.
• The method facilitates reliable analysis of structural changes, relevant for conditions like Alzheimer's disease.