SynthStrip

SynthStrip performs learning-based brain extraction (skull-stripping) from magnetic resonance imaging (MRI) acquisitions to produce brain masks for downstream neuroimage analysis.

Key Features:

• Learning-based brain extraction: applies a rapid learning-based method to generate brain masks from MRI images.
• Synthetic training dataset: trained exclusively on an entirely synthetic dataset generated from anatomical segmentations.
• Diverse simulated variability: the synthetic dataset encompasses a broad range of anatomies, intensity distributions, and artifacts beyond the realistic scope of medical images.
• Contrast and acquisition agnostic: generalizes across T1-weighted (T1w), fast spin-echo (FSE), and other clinical MRI contrasts without target contrast-specific training data.
• Resolution robustness: supports near-isotropic and thick-slice stacks (e.g., FSE) across varying spatial resolutions.
• Single-model generalization: achieves robust skull-stripping performance across multiple acquisitions and populations with a single trained model.
• Improved accuracy: demonstrates accuracy improvements over popular skull-stripping baselines.
• Population range: validated across subject populations from newborns to adults.

Scientific Applications:

• Preprocessing for neuroimage analysis: provides brain masks for downstream workflows in neuroimaging studies.
• Clinical MRI processing: applicable to clinical imaging types including thick-slice FSE and non-T1w contrasts encountered in hospital settings.
• Cross-population studies: enables brain extraction across age ranges and diverse cohorts for comparative analyses.

Methodology:

Leveraging anatomical segmentations, SynthStrip synthesizes training images with varied anatomies, intensity distributions, and artifacts and trains a learning-based model on these synthetic images to produce a single generalizable brain-extraction model.