Segmentation of MRI brain scans using spatial constraints and 3D features.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate brain tissue segmentation is crucial for neurological disorder diagnosis and research.
• Existing unsupervised methods often struggle with precision and robustness across diverse datasets.
• Magnetic Resonance Imaging (MRI) is a primary modality for non-invasive brain imaging.

Purpose of the Study:

• To introduce Gardens2, a novel unsupervised algorithm for segmenting brain tissues in MRI.
• To evaluate Gardens2's performance against established and new unsupervised segmentation techniques.
• To demonstrate the efficacy of 3D features and atlas information in enhancing segmentation accuracy.

Main Methods:

• Gardens2 employs a clustering approach to classify MRI voxels into cerebrospinal fluid (CSF), gray matter (GM), and white matter (WM).
• The algorithm utilizes a 3D feature descriptor and an overlapping criterion with prior atlas information.
• Segmentation masks are generated per class for detailed brain tissue parcellation.

Main Results:

• Gardens2 achieved superior segmentation performance on two out of three neuroimaging datasets (BrainWeb, IBSR18, IBSR20).
• The algorithm demonstrated competitive results when performance was evaluated on a per-class basis.
• Compared to eleven other unsupervised methods, Gardens2 showed improved overall accuracy.

Conclusions:

• Gardens2 represents a significant advancement in unsupervised brain tissue segmentation using MRI.
• The integration of 3D features and adjusted atlas templates enhances segmentation reliability.
• This method offers a robust tool for neuroimaging research and clinical applications.