Markov random field segmentation of brain MR images

Area of Science:

• Medical imaging analysis
• Computational neuroscience
• Biomedical image processing

Background:

• Accurate segmentation of brain magnetic resonance (MR) images is crucial for quantitative analysis.
• Existing methods may struggle with non-parametric intensity distributions, spatial correlations, and signal inhomogeneities inherent in MR data.

Purpose of the Study:

• To develop and evaluate a fully-automatic three-dimensional (3-D) segmentation technique for brain MR images.
• To incorporate key MR image features into the segmentation algorithm for improved accuracy.

Main Methods:

• Utilized Markov random fields (MRF) to model non-parametric tissue intensity distributions, neighborhood correlations, and signal inhomogeneities.
• Implemented simulated annealing and iterated conditional modes (ICM) for the segmentation process.
• Performed quantitative analysis using simulated and real MR images, assessing the impact of noise, inhomogeneity, smoothing, and structure thickness.

Main Results:

• The algorithm successfully segmented single-echo MR images into distinct tissue classes: gray matter, white matter, cerebrospinal fluid, scalp-bone, and background.
• Demonstrated robust performance across various noise levels and signal inhomogeneity conditions.
• Quantitative analysis provided insights into the algorithm's sensitivity to imaging parameters.

Conclusions:

• The developed fully-automatic 3-D MR image segmentation technique effectively utilizes MRF to capture essential image characteristics.
• The method offers accurate classification of brain tissues, proving valuable for neuroimaging research and clinical applications.
• The quantitative evaluation confirms the algorithm's reliability and performance under challenging imaging conditions.