Probabilistic segmentation of brain tissue in MR imaging.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate segmentation of brain structures in cranial Magnetic Resonance (MR) imaging is crucial for diagnosing and monitoring neurological conditions.
• Existing methods may struggle with differentiating specific tissue types, including white matter lesions.

Purpose of the Study:

• To develop and evaluate a new probabilistic segmentation method for five distinct brain tissue types in cranial MR imaging.
• To determine the optimal combination of MR imaging sequences for accurate brain tissue segmentation.

Main Methods:

• A K-Nearest Neighbor classification technique was employed, utilizing spatial information and voxel intensities from T1-weighted (T1-w), inversion recovery (IR), proton density-weighted (PD), T2-weighted (T2-w), and fluid attenuation inversion recovery (FLAIR) scans.
• The method generates probability maps for each tissue type, enabling binary segmentation through thresholding.
• Quantitative evaluation was performed using Similarity Index (SI) and Probabilistic SI (PSI) metrics.

Main Results:

• The study found that T1-w, PD, and T2-w scans did not significantly enhance segmentation performance when added to other sequences.
• The optimal combination for segmenting the five brain tissue types was identified as the integration of IR and FLAIR scans.
• Evaluation against a gold standard demonstrated excellent agreement, with SI-values exceeding 0.8 and PSI-values exceeding 0.7 for all tissues.

Conclusions:

• The developed probabilistic segmentation method provides accurate differentiation of white matter, gray matter, cerebro-spinal fluid, ventricles, and white matter lesions.
• The combination of IR and FLAIR sequences is highly effective and recommended for robust brain tissue segmentation in MR imaging.
• The method shows excellent agreement with gold standards, highlighting its potential clinical utility.