pydcm.transforms — authoritative references & precision

There is no single gold-standard transforms library in Python. Medical-image ML preprocessing is fragmented, and two numeric families genuinely disagree:

• scipy.ndimage family (B-spline resampling, filtering, connected components) — the de-facto numeric reference; nnU-Net and skimage build on it.
• torch family (grid_sample) — a separate, incompatible interpolation convention; MONAI Spacing/Resize use it.

"Correct" for a preprocessing op therefore means "matches the implementation the target framework actually used to train the model" — it is defined per-framework, not by math. pydicom is not in this picture: it does DICOM-defined pixel transforms only (rescale / modality LUT / VOI LUT), never spatial/ML preprocessing. Validate against the op's own upstream below — never against another reproduction (custom_ops, pydcm) — those share lineage and would validate circularly.

To avoid mixing families, import a preset (from pydcm.transforms import nnunet as T or monai as T); it binds the divergent ops and shares the rest.

Per-op reference

op	authoritative reference	precision vs reference
resample_cubic	scipy.ndimage / skimage.resize(order=3)	bit-exact
resample_separate_z	scipy.ndimage (per-slice cubic + nearest-z, nnU-Net)	bit-exact
resample_nearest	scipy.ndimage order=0	bit-exact
resample_grid_sample	torch.nn.functional.grid_sample (align_corners=False)	≤1 fp32 ULP¹
resample_to_spacing, resize	MONAI Spacing/Resize (half-pixel)	≤1 fp32 ULP¹
normalize_ct	nnU-Net CTNormalization (clip + fixed mean/std)	bit-exact
normalize_zscore	MONAI NormalizeIntensity (self-computed stats)	bit-exact²
scale_intensity_range	MONAI ScaleIntensityRange	tight tolerance
scale_intensity_range_percentiles	np.percentile (linear) + MONAI	tight tolerance
adjust_contrast	MONAI AdjustContrast	~1e-5 (powf)
crop, pad, crop_foreground, flip	numpy / MONAI definitions (exact index)	exact
center_crop, spatial_pad, divisible_pad	MONAI CenterSpatialCrop/SpatialPad/DivisiblePad	exact
rotate90	np.rot90 (+ world-coord preservation)	exact
gaussian_smooth	MONAI GaussianSmooth (reflect, separable)	tight tolerance
argmax	np.argmax over channels	exact
connected_components	scipy.ndimage label	exact
keep_largest_connected_component, fill_holes, remove_small_objects, as_discrete	MONAI post-transforms	exact
sliding_window_positions	MONAI dense_patch_slices / _get_scan_interval	exact
gaussian_importance_map(convention='nnunet')	nnU-Net V2 get_gaussian (scipy gaussian_filter)	bit-exact
gaussian_importance_map(convention='monai')	MONAI compute_importance_map (sampled exp)	≤1 fp32 ULP¹

¹ Not bit-exact: the reference's last-bit rounding is not reproducible from any closed form (torch's affine_grid SIMD linspace; expf). Numerically identical otherwise. ² np.std ddof=0 (population), matching MONAI's default.

Presets — which op a preset binds

preset	resample	gaussian_importance_map / sliding_window_inference
transforms.nnunet	scipy cubic spline (labels → nearest)	convention='nnunet'
transforms.monai	torch grid_sample (labels → nearest)	convention='monai'

Every other op is identical in both presets and in the flat module. For nnU-Net's anisotropic thick-slice path, call resample_separate_z explicitly.