SPINEPS v1.0.9 nnU-Net V2 PlainConvUNet (T2w_Segmentor_2.0, fold 0,
L=7, patch [64,256,256], target spacing
[1.65, 0.75, 0.75] mm, anisotropic L0 kernel
[1,3,3], 14-class subregion labels + background).
All compute (WASM + WebGPU) runs in a dedicated Web Worker —
page stays responsive during inference. Reorients input to PIR
to match SPINEPS's expected orientation.
How to use:
① Init WASM + model → ② Upload sagittal T2 MR NIfTI (.nii / .nii.gz)
→ ③ Run → ④ Download.
Inputs with anisotropy ratio > 3× go through host-side separate_z
per-slice 2D B-spline + nearest-Z; near-iso (ratio ≤ 3×) uses GPU 3D B-spline.
SPINEPS T2w target ratio is 1.65/0.75 = 2.2× (under the threshold).
Weights are cached in IndexedDB on first load (~178 MB); subsequent visits
are offline-ready.