Skip to content

README.md

Latest commit

 

History

History
73 lines (51 loc) · 2.38 KB

README.md

File metadata and controls

73 lines (51 loc) · 2.38 KB
Raw

Brainchomp

Brainchomp is a browser-only rodent brain-extraction app derived from Brainchop. Imaging data stays on the user's device.

Live app: neuroneural.net/brainchomp

Model and outputs

The single Skull-strip model runs the rodent MeshNet and retains its largest connected foreground component. One inference produces both downloadable outputs:

  • a skull-stripped image that preserves the input intensities and sets non-brain voxels to zero;
  • a post-processed binary brain mask.

The Mask toggle switches the overlay without rerunning inference. The brain uses the same copper overlay as Brainchop's MindGrab result; the mask uses a binary label overlay. The Save menu always offers both files.

Input contract

This version intentionally does not conform or resample inputs. NIfTI volumes must already be 256 × 256 × 256. Model preprocessing uses the 2nd and 98th intensity percentiles with a 1e-3 denominator guard and clipping to [0, 1], matching MindGrab in brainchop-cli.

Two bundled volumes can be selected from the toolbar:

  • MA_F_15M_20_image.nii.gz — rodent MRI
  • camriMouse_sub-012_ses-1_t2w_image.nii.gz — CAMRI mouse T2-weighted MRI

Both supplied files are min/max-quantized 256³ uint8 NIfTIs with 0.1 mm isotropic voxel spacing. Their spatial geometry is preserved from the source volumes.

Model and runtimes

The single checkpoint is a 25-block MeshNet with 16 hidden channels, non-affine per-channel GroupNorm, GELU, and two output classes. Foreground is class 1. The app prefers the generated fp16 WebGPU runner and falls back to the native WebGL2 runner. A TF.js model is retained as a final compatibility fallback.

The WebGPU export uses the low-memory Brainchop pipeline: fp16 activation storage, fp32-safe GroupNorm conditioning, and fused two-class argmax. The largest individual activation buffer is 512 MiB.

Development

npm ci
npm run dev

Build and verification:

npm run build
node tests/webgl2_gate.mjs chromium

Append ?backend=webgl2 to the app URL to exercise the native WebGL2 fallback on a browser that also supports WebGPU.

The model conversion inputs and reproducibility manifest live in temporaryRodentModel/; browser-ready assets live in public/models/rodent/.

License

Brainchomp inherits Brainchop's MIT license.