Fully automated classification of HARDI in vivo data using a support vector machine.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Engineering

Background:

• High angular resolution diffusion imaging (HARDI) provides detailed information about brain microstructure.
• Accurate segmentation of HARDI data is crucial for quantitative analysis and clinical applications.
• Current segmentation methods often rely on T1-weighted scans and manual intervention, limiting automation.

Purpose of the Study:

• To develop and validate a model-free approach for classifying in vivo HARDI data.
• To enable fully automatic segmentation of HARDI images using machine learning.
• To improve the performance of subsequent analyses like fibre tracking.

Main Methods:

• Utilized a Support Vector Machine (SVM) algorithm from supervised statistical learning.
• Employed spherical harmonic decomposition of HARDI signals to create rotation-invariant feature vectors.
• Trained the SVM to classify six distinct image components: grey matter, white matter fibre bundles (parallel and crossing), partial volumes, noise, and cerebrospinal fluid.

Main Results:

• The SVM model was systematically tested on simulated data and applied to six in vivo datasets.
• Achieved robust and accurate segmentation results across multiple in vivo datasets.
• Demonstrated the capability of fully automatic HARDI data segmentation without T1 MPRAGE scans or expert input.

Conclusions:

• The developed data-driven, model-free approach offers a robust solution for automatic HARDI data segmentation.
• The segmentation results can serve as a priori knowledge to enhance fibre tracking performance.
• This method has potential for broader clinical and diagnostic applications in diffusion-weighted imaging (DWI).