[1, 3, 4, 6, 6],
anisotropic deepest stride (1, 2, 2).
·
worker product path v2
·
legacy full-volume page
·
DentalSegmentator parity/debug
h = lrelu(norm2(conv2(lrelu(norm1(conv1(x))))) + skip(x)))
where skip is a 1×1×1 conv + InstanceNorm projection when
channels or spatial dimensions change, or identity otherwise. This is the
"ResEncUNet" variant in nnU-Net V2 — served by our resenc_model.c
runtime rather than unet_model.c.