MONAI VISTA3D, Phase 1 auto branch (no point prompts yet). Whole-body CT
segmentation with a CLASS prompt — user picks a class name; the model
returns a binary mask for that organ/structure. 196 M params, SegResNet
encoder + dual decoder + class-embedding matmul head.
·
custom_ops WebGPU runtime: WGSL conv3d / convT3d(opad) / IN3d / gemm.
✓ Node M1 verification (2026-04-18):
per-patch rel_rms 1.5e-6 vs PyTorch,
213×213×163 sliding window rel_rms 4.4e-7, binary agree
100.0000% vs MONAI SlidingWindowInferer.
Matches MPS baseline throughput (~5 s/patch on M1).
⚠ Demo scope: ships a pre-canonicalized CT volume (already
ScaleIntensityRanged + Spacingd 1.5 mm, shape 213×213×163). Upload-your-own
NIfTI flow and the 117-class EVERYTHING_PROMPT one-shot mode are future work.
· Download: ~786 MB of weights + 28 MB canonical volume.