3D volumetric nucleus INSTANCE segmentation. A 3D U-Net predicts a 3-class volume (bg / interior / boundary); instances are recovered scipy-only by labelling 3D interior blobs and expanding to the foreground Voronoi region. Trained on NucMM-Z train tiles (per-volume percentile norm). predict(image (C,i,j,k)) -> (i,j,k) int32 instance ids on the native grid.