Automated voxel-by-voxel tissue classification for hippocampal segmentation: methods and validation

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• The hippocampus is a key biomarker in Alzheimer's disease (AD) and other neurological disorders.
• Accurate segmentation of the hippocampus in structural MRI is essential for disease diagnosis and research.
• Current segmentation methods may lack accuracy and reproducibility.

Purpose of the Study:

• To develop and validate a fully automated pattern recognition system for accurate and reproducible hippocampus segmentation in structural MRI.
• To improve the performance of existing segmentation techniques for Alzheimer's disease research.

Main Methods:

• A three-level processing system involving linear registration, feature extraction (315 features), and Random Forest voxel classification.
• An adaptive learning method using Pearson's correlation coefficient was developed to enhance classification performance.
• Validation on a mixed cohort of 56 T1-weighted MRI images and an independent dataset of 100 T1-weighted MRI images.

Main Results:

• The automated system achieved a Dice similarity index of 0.81 ± 0.03 on the validation cohort.
• Segmentation results demonstrated comparable performance to state-of-the-art approaches.
• The system significantly outperformed FreeSurfer on an independent dataset.

Conclusions:

• The developed automated system provides accurate and reproducible hippocampus segmentation from structural MRI.
• This method holds promise for advancing Alzheimer's disease research and the study of other neurological conditions.
• The system offers a valuable tool for neuroimaging analysis, potentially improving diagnostic capabilities.