Dataset852 nnU-Net V2 PlainConvUNet (fold 0,
[64,64,64] patch / 3.0 mm iso spacing,
51-class TotalSegMRI labels). All compute (WASM + WebGPU) runs in a
dedicated Web Worker — page stays responsive during inference.
How to use:
① Init WASM + model → ② Upload MR NIfTI (.nii / .nii.gz) → ③ Run → ④ Download.
Anisotropic inputs (max/min spacing > 3×) go through host-side separate_z
per-slice 2D B-spline + nearest-Z (matches nnU-Net exactly). Near-iso inputs use
the faster GPU 3D B-spline path.
Model weights are cached in IndexedDB on first load (66 MB); subsequent visits
are offline-ready.