Brain Extraction
Brain extraction removes nonbrain tissue from head images.
Scalp removal
Most neuroimaging pipelines remove signal from the scalp, skull, and other nonbrain tissues. Brain extraction can improve registration and restrict subsequent analyses to the brain.
FSL BET established a widely used conventional approach. SynthStrip later introduced robust, modality-agnostic extraction using deep learning. SynthStrip benefits from NVIDIA GPU acceleration but can be slow when restricted to a CPU.
Brainchop provides a smaller model that runs through WebGPU on supported Apple, AMD, Intel, and NVIDIA hardware. A browser demo is available. The lighter mindgrab model is optimized for CPU execution, which can be useful when a pipeline must reserve GPU resources for other tasks. niimath includes the CPU-based mindgrab model.
Links
- Iglesias et al. (2011) evaluate established methods and introduce ROBEX.
- Hoopes et al. (2022) describe SynthStrip.
- Fani et al. (2026) introduce MindGrab, a lightweight Brainchop model trained with SynthStrip-derived images.
- Smith (2002) describes BET.
- FSL BET software (institutional license).
- brainchop mindgrab source code (permissive MIT license).
- SynthStrip code and training data.