Feature Selection Based on Machine Learning in MRIs for Hippocampal Segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Neuroscience

Background:

• Neurodegenerative diseases often involve brain structural changes detectable by magnetic resonance imaging (MRI).
• The hippocampus is a key biomarker for Alzheimer disease and other neurological/psychiatric conditions, necessitating precise delineation.
• Automated voxel-based methods require effective feature selection for accurate hippocampal segmentation.

Purpose of the Study:

• To compare four distinct feature selection techniques for voxel-based hippocampal segmentation.
• To evaluate the performance of these methods against manual segmentation and a standard tool (FreeSurfer).
• To identify an optimal feature selection strategy for robust and reproducible hippocampal structure delineation.

Main Methods:

• Extraction of 315 local features per voxel from T1-weighted brain MRIs.
• Comparison of four feature selection methods: Kolmogorov-Smirnov test (filter), sequential forward selection (wrapper), sequential backward elimination (wrapper), and Random Forest Classifier (embedded).
• Testing on an independent dataset of 25 subjects, with segmentations compared to manual reference labeling.

Main Results:

• Sequential backward elimination achieved state-of-the-art performance using only 23 features per voxel.
• The performance was comparable to the widely used FreeSurfer tool.
• This indicates high efficiency and accuracy in feature selection for hippocampal segmentation.

Conclusions:

• Sequential backward elimination is a highly effective method for selecting features in automated hippocampal segmentation.
• This approach offers a robust and reproducible alternative to existing methods for neuroimaging analysis.
• The findings support the use of optimized feature selection for improved diagnostic support in neurodegenerative diseases.