Segmentation of brain tissues using a 3-D multi-layer hidden Markov model.

Area of Science:

• Medical Image Analysis
• Computational Neuroscience
• Machine Learning

Background:

• Bias field inhomogeneity and noise degrade the quality of brain MRI scans.
• Accurate segmentation of brain tissues (Gray Matter, White Matter, Cerebrospinal Fluid) is crucial for neurological studies.
• Existing algorithms often struggle with artifacts and complex spatial relationships within brain images.

Purpose of the Study:

• To develop a robust brain segmentation algorithm that addresses bias field inhomogeneity and noise.
• To leverage domain knowledge and spatial information for improved segmentation accuracy.
• To propose and evaluate a novel multi-layer Hidden Markov model for 3D medical image segmentation.

Main Methods:

• A multi-layer Hidden Markov model (HMM) was proposed, integrating domain knowledge and spatial information.
• The first layer employed a 1-D HMM to classify image slices into three categories (GM, GM-WM, GM-WM-CSF).
• The second layer utilized another 1-D HMM for slice segmentation, processing slices as concatenated row vectors.

Main Results:

• The multi-layer HMM demonstrated significant potential for segmenting 3D medical images with noise and field inhomogeneity.
• Evaluation on three public datasets (5492 images) validated the method's effectiveness.
• Specifically, the IBSR_18 dataset showed improvements in White Matter and Gray Matter segmentation by 0.026 and 0.04 (Dice coefficient), respectively.

Conclusions:

• The proposed multi-layer Hidden Markov model offers a promising approach for accurate brain MRI segmentation.
• The method effectively handles challenges like noise and bias field inhomogeneity.
• This technique has significant implications for quantitative analysis in neuroimaging research.