Brain MAPS: an automated, accurate and robust brain extraction technique using a template library.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Automated brain extraction is vital for large-scale neuroimage analysis, reducing labor compared to manual methods.
• Accuracy and robustness of automated techniques are critical to minimize time-consuming manual corrections.
• Existing methods like Brain Extraction Tool (BET), Brain Surface Extractor (BSE), and Hybrid Watershed Algorithm (HWA) have varying performance.

Purpose of the Study:

• To compare the accuracy and robustness of four automated brain extraction methods: BET, BSE, HWA, and a novel Multi-Atlas Propagation and Segmentation (MAPS) technique.
• To evaluate method performance on T1-weighted MR images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database at 1.5T and 3T.
• To identify the most accurate and reliable automated brain extraction method for neuroimaging studies, especially those involving atrophy.

Main Methods:

• Applied four automated brain extraction algorithms (BET, BSE, HWA, MAPS) to 682 1.5T and 157 3T T1-weighted MR images from the ADNI database.
• Utilized semi-automated brain segmentations with manual editing as the gold standard for comparison.
• Quantified accuracy using the Jaccard index and assessed tissue inclusion via false negative rates.

Main Results:

• MAPS demonstrated a significantly higher median Jaccard index than HWA, BET, and BSE across both 1.5T and 3T scans (p<0.05).
• MAPS exhibited lower variability (smaller Jaccard index range) compared to HWA, BET, and BSE (p<0.05).
• HWA and MAPS showed superior performance in including all brain tissues, with low median false negative rates (≤0.010% for 1.5T, ≤0.019% for 3T).
• MAPS maintained consistent performance across 1.5T and 3T, unlike BET, BSE, and HWA which performed better at 1.5T.
• Diagnostic group had a minimal effect on the performance of all tested methods.

Conclusions:

• MAPS provides superior accuracy and lower variability for automated whole-brain extraction in MR images, including those with atrophy, compared to BET, BSE, and HWA.
• MAPS is a robust and reliable automated method suitable for large-scale neuroimaging studies utilizing diverse MR field strengths.
• The findings support the use of MAPS for efficient and accurate pre-processing in neuroimage analysis, particularly within the context of Alzheimer's disease research.