Abdominal CT 14-class segmentation — all compute (WASM + WebGPU) runs in a
dedicated Web Worker so the page stays fully responsive during the ~120 s
inference. MONAI pipeline: NormalizeIntensity(nonzero=True) → sliding window
(96³, overlap=0.5, constant).
✓ MONAI strict conformance:100% argmax agreement vs MONAI SlidingWindowInferer
(147 patches, 24 M voxels, BTCV abdominal CT). Logits rel_rms = 2.57e-6.
Chrome WebGPU ≈ Node.js (~120 s on M1, 823 ms/patch).
⚠ Download size: ~253 MB of SwinUNETR BTCV weights will be fetched.