Robust brain extraction across datasets and comparison with publicly available methods.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Automatic brain extraction (skull stripping) is crucial for neuroimaging analysis.
• Existing methods often lack robustness across different MRI acquisition conditions and populations.
• This limitation impacts the reliability of downstream neuroimage processing pipelines.

Purpose of the Study:

• To introduce ROBEX, a robust, learning-based system for automatic brain extraction.
• To evaluate ROBEX's performance against established skull stripping methods.
• To demonstrate ROBEX's generalizability across diverse neuroimaging datasets.

Main Methods:

• ROBEX integrates a Random Forest classifier (discriminative model) for boundary detection and a point distribution model (generative model).
• It utilizes graph cuts for refining segmentation contours based on model likelihoods.
• The system was trained on a proprietary dataset and validated on multiple public datasets.

Main Results:

• ROBEX demonstrated significantly improved performance compared to six other popular skull stripping methods (BET, BSE, FreeSurfer, AFNI, BridgeBurner, GCUT).
• Robustness was observed across three diverse public datasets (IBSR, LPBA40, OASIS) with varying hardware and populations.
• Performance improvements were consistent across almost all method/dataset combinations.

Conclusions:

• ROBEX offers a robust and accurate solution for automated brain extraction in MRI.
• Its learning-based approach enhances generalizability, addressing limitations of previous methods.
• ROBEX represents a significant advancement for reliable neuroimage analysis pipelines.