MindGrab for BrainChop: Fast and Accurate Skull Stripping for Command Line and Browser

Deployment complexity and specialized hardware requirements hinder the adoption of deep learning models in neuroimaging. We present MindGrab, a lightweight, fully convolutional model for volumetric skull stripping across all imaging modalities. MindGrab’s architecture is designed from first principles using a spectral interpretation of dilated convolutions, and demonstrates state-of-the-art performance (mean Dice score across datasets and modalities: 95.9 with SD 1.6), with up to 40-fold speedups and substantially lower memory demands compared to established methods. Its minimal footprint allows for fast, full-volume processing in resource-constrained environments, including direct in-browser execution. MindGrab is delivered via the BrainChop platform as both a simple command-line tool (pip install brainchop) and a zero-installation web application (brainchop.org). By removing traditional deployment barriers without sacrificing accuracy, MindGrab makes state-of-the-art neuroimaging analysis broadly accessible.

💡 Research Summary

The paper addresses a critical bottleneck in neuroimaging: the gap between high‑accuracy deep‑learning skull‑stripping methods and their practical deployment in clinical or low‑resource research settings. While modern models such as SynthStrip achieve impressive robustness by training on synthetic data, they remain heavyweight, requiring substantial GPU memory, complex software stacks, and often a cloud environment that raises privacy concerns. To bridge this gap, the authors introduce MindGrab, a lightweight fully‑convolutional network specifically engineered for volumetric skull stripping across all major imaging modalities (T1w, T2w, PDw, MRA, DWI, CT, PET, etc.).

The core technical contribution is a spectral reinterpretation of dilated convolutions. By viewing dilation as a frequency‑domain filter that replicates a small kernel’s response across k‑space, the authors design a dilation schedule that directly controls which spatial frequencies the network can sense. They adopt a decreasing dilation pattern (16 → 8 → 4 → 2 → 1) across five blocks, each block containing five 3×3×3 convolutions with 15 channels, followed by a final 1×1×1 projection. All layers are isometric, bias‑free, and use GeLU activation with instance normalization. This results in a 26‑layer network with roughly 0.2 M parameters—about 95 % fewer than typical U‑Net‑based skull‑stripping models—yet it retains a receptive field large enough to process a full 256³ volume in a single pass. Memory efficiency is further enhanced by storing only one activation map during inference, enabling execution on consumer‑grade CPUs and even in‑browser via WebGL/WebGPU.

Training relies entirely on synthetic data generated by the Wirehead pipeline, which builds on SynthSeg to produce 250 k volumetric samples from 171 label maps (39 anatomical structures each). Randomized spatial deformations, intensity scaling, resolution changes, and realistic artifacts (blur, noise) are applied, ensuring domain‑agnostic feature learning. The model is optimized with Adam, a soft Dice loss, and a One‑Cycle learning‑rate schedule over 50 cycles, batch size = 1.

Evaluation uses the multimodal benchmark originally assembled for SynthStrip, comprising 606 adult scans from eight public datasets, covering MRI (T1, T2, PD, MRA, DWI), quantitative T1, ASL, EPI, CT, PET, and clinical thick‑slice protocols. Ground truth is a “silver standard” mask generated by averaging multiple automatic methods. MindGrab achieves a mean Dice of 95.9 ± 1.6 % across all modalities, significantly outperforming classical tools (ROBEX, BET) and matching SynthStrip within a 3 % margin. Precision is consistently higher than SynthStrip (favoring fewer false positives), while recall is slightly lower, reflecting a conservative segmentation bias. Speed tests show up to 40× acceleration compared with SynthStrip, with inference times of ~0.3 s on a high‑end GPU and ~2 s on a standard CPU, and a memory footprint reduced by more than 70 %.

Deployment is realized through the BrainChop platform. Users can install a command‑line tool via pip install brainchop or access a zero‑installation web application at brainchop.org, where the model runs entirely client‑side, preserving patient privacy and eliminating the need for CUDA drivers or complex dependency management.

In summary, MindGrab combines a principled spectral design of dilated convolutions with exhaustive synthetic training to deliver state‑of‑the‑art skull stripping that is fast, memory‑light, and trivially deployable on both local machines and web browsers. This work paves the way for broader adoption of deep‑learning neuroimaging tools in routine clinical workflows and resource‑constrained research environments.