Preprocessing

pydcm.transforms ports the spatial and intensity operations a segmentation pipeline needs, with the numerics pinned to the framework you train against.

Why two precision tiers

There is no single gold-standard preprocessing library in Python. The classic B-spline family and the deep-learning (grid-sample) family use genuinely different interpolation conventions, so "correct" is defined per framework. pydcm matches each one exactly. Details: Transforms — precision & references.

• Tier 1 — resample_cubic, resample_separate_z, resample_nearest are bit-exact for the classic B-spline convention.
• Tier 2 — resample_grid_sample matches the deep-learning trilinear convention to ≤ 1 float32 ULP.

Resample

resample_to_spacing resamples a Volume to an axis-aligned grid at a target mm spacing (the common "make it isotropic" step):

import pydcm
from pydcm import transforms as T

vol = pydcm.load_series("ct_series/")
iso = T.resample_to_spacing(vol, (1.0, 1.0, 1.0), interp="linear")
seg = T.resample_to_spacing(label_vol, (1.0, 1.0, 1.0), is_label=True)  # nearest, no new labels

When you already know the output shape, choose the family that matches your model:

out = T.resample_cubic(vol, (128, 256, 256))        # Tier 1 — bit-exact B-spline order-3
out = T.resample_separate_z(vol, (128, 256, 256))   # anisotropic separate-z behaviour
out = T.resample_grid_sample(vol, (128, 256, 256))  # Tier 2 — deep-learning trilinear convention

Intensity normalization

# CT normalization: clip to a window, then z-score with FIXED dataset-level
# mean/std (computed during planning):
ct = T.normalize_ct(vol, clip_lo=-1000, clip_hi=400, mean=-380.0, std=320.0)

z = T.normalize_zscore(vol, nonzero=True)            # per-volume, ignore background zeros

Sliding-window inference

logits = T.sliding_window_inference(
    vol.pixels,
    roi_size=(96, 96, 96),
    predictor=model,          # a callable patch → logits
    overlap=0.25,
    convention="nnunet",      # or "monai" — picks the gaussian importance map
)

Label post-processing

seg = T.argmax(logits, vol.affine)                   # logits [C, D, H, W] → labelled Volume
seg = T.keep_largest_connected_component(seg, per_class=True)
seg = T.remove_small_objects(seg, min_size=64)
seg = T.fill_holes(seg)
onehot = T.one_hot(seg, num_classes=3)               # [C, D, H, W]

Every op accepts a Volume (or a plain array where noted) and preserves the geometry, so you can resample, infer, post-process, and write the result straight back to a SEG (see Segmentations) without losing the affine.