Self-Trained Supervised Segmentation of Subcortical Brain Structures Using Multispectral Magnetic Resonance Images.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Image Analysis

Background:

• Supervised classification methods for brain structure segmentation require extensive manual training data.
• Operator intervention for training data selection limits the widespread adoption of these methods.

Purpose of the Study:

• To investigate the feasibility of automatically training supervised methods for segmenting subcortical brain structures.
• To evaluate atlas-guided training for k-nearest neighbor (kNN) and principal component discriminant analysis (PCDA) classifiers.

Main Methods:

• Automated training data selection using probabilistic atlas registration.
• Utilized voxel intensities and spatial coordinates for kNN and PCDA classifiers.
• Evaluated on 20 multispectral magnetic resonance imaging datasets.

Main Results:

• Atlas-guided training effectively creates representative and reliable datasets automatically.
• Achieved successful segmentation of caudate, thalamus, pallidum, and putamen.
• Demonstrated the potential for fully automated supervised segmentation of brain images.

Conclusions:

• Automated atlas-guided training enables supervised methods for brain MRI segmentation without user interaction.
• This approach enhances the feasibility and efficiency of neuroimaging analysis.
• Paves the way for wider application of automated segmentation techniques.