Two ResEncUNet L=7 models + host-side iterative anchor labeling for per-vertebra /
per-disc instance segmentation on lumbar-and-cervical MRI.
[256,256,96] patch / 1.0 mm iso. All compute in C/WGSL.
Pipeline: NIfTI → crop + z-score → separate_z →
step1 (9 regions)
→
CC chain (anchor labeling)
→
step2 Cin=2 (11 regions)
→ inverse separate_z (logits) → segmentation NIfTI.
Weights (~1.1 GB) cached in IndexedDB on first load.