neu_proc.ops.kernels¶

The one place a physical length becomes voxels.

Parameters are declared in nanometres and functions take voxels; the conversion lives only here, for two reasons:

A halo must be derived from the converted parameter. A Gaussian whose halo assumes truncate=3.0 while scipy uses 4.0 is wrong in a shell around every block seam, which shows as faint block edges rather than as an error.
A voxel count means nothing without a Frame: sigma_vox=2 is different smoothing at level 0 and level 2, and along z and x on an anisotropic volume.

Nothing here imports scipy.

Attributes¶

GAUSSIAN_TRUNCATE

Functions¶

`ball`(radius_vox, *[, rank]) → numpy.ndarray	A boolean footprint of this per-axis radius in voxels: an ellipsoid when the radii
`box`(size_vox, *[, rank]) → tuple[int, ...]	An odd per-axis footprint size in whole voxels, at least 1. Odd, because an even
`gaussian_halo`(sigma_vox, *[, truncate, rank]) → tuple[int, ...]	The read margin a Gaussian of this sigma needs, per axis, in whole voxels.
`per_axis`(value[, rank]) → tuple[float, ...]	A scalar or per-axis sequence as a length-`rank` tuple, zyx; a scalar is repeated.
`to_voxels`(length_nm, voxel_size_nm, *[, rank]) → tuple[float, ...]	A physical length, per axis, in voxels of `voxel_size_nm`. Not rounded: a sigma
`volume_to_voxels`(volume_nm3, voxel_size_nm) → float	A volume in nm^3 as a voxel count, dividing by the product of the voxel sizes (a

Module Contents¶

neu_proc.ops.kernels.ball(radius_vox: float | Sequence[float], *, rank: int = 3) → numpy.ndarray¶: A boolean footprint of this per-axis radius in voxels: an ellipsoid when the radii differ, which is what a physical ball becomes on an anisotropic volume.

neu_proc.ops.kernels.box(size_vox: float | Sequence[float], *, rank: int = 3) → tuple[int, ...]¶: An odd per-axis footprint size in whole voxels, at least 1. Odd, because an even window shifts a centred filter’s result by half a voxel.

neu_proc.ops.kernels.gaussian_halo(sigma_vox: float | Sequence[float], *, truncate: float = GAUSSIAN_TRUNCATE, rank: int = 3) → tuple[int, ...]¶

The read margin a Gaussian of this sigma needs, per axis, in whole voxels.

ceil(truncate * sigma), scipy’s kernel radius. Pass the same voxel sigma and truncate the filter will get.

neu_proc.ops.kernels.per_axis(value: float | Sequence[float], rank: int = 3) → tuple[float, ...]¶: A scalar or per-axis sequence as a length-rank tuple, zyx; a scalar is repeated.

neu_proc.ops.kernels.to_voxels(length_nm: float | Sequence[float], voxel_size_nm: Sequence[float], *, rank: int = 3) → tuple[float, ...]¶: A physical length, per axis, in voxels of voxel_size_nm. Not rounded: a sigma stays fractional, while a footprint radius is rounded by its caller.

neu_proc.ops.kernels.volume_to_voxels(volume_nm3: float, voxel_size_nm) → float¶: A volume in nm^3 as a voxel count, dividing by the product of the voxel sizes (a 40x8x8 nm voxel is 2560 nm^3, so 1 um^3 is 390 voxels). Not rounded; the caller picks ceil or floor.

neu_proc.ops.kernels.GAUSSIAN_TRUNCATE = 4.0¶