wenmin-wu/ds-skills/anisotropic-spacing-trilinear-resample

Overview

CT volumes are routinely anisotropic: 0.7mm in-plane and 5mm between slices is normal. If you F.interpolate a (60, 512, 512) volume directly to (96, 256, 256), you stretch the Z axis 1.6x in voxel space but the physical distance per slice is already 7x the in-plane spacing — the resulting volume looks correct in tensor shape but is geometrically wrong, and the network learns the wrong aspect ratio. The fix is to first scale the slice count by dz / dy (the spacing ratio), then resample to the target shape. This delivers an anatomically isotropic input regardless of the source acquisition.

Quick Start

import torch
import torch.nn.functional as F
def resample_to_shape(image, spacing, target_hw=256, target_d=96):
    dz, dy, dx = spacing             # physical mm/voxel
    d, h, w = image.shape
    # 1. Correct Z count by physical-spacing ratio (anisotropy fix)
    d_iso = int(dz / dy * d * 0.5)   # 0.5 = empirically tuned compression
    # 2. Scale all axes to in-plane target
    scale = target_hw / h
    d_out = int(scale * d_iso)
    h_out = w_out = int(scale * h)
    # 3. Final resample to fixed network input shape
    image = F.interpolate(image[None, None],
                          size=(d_out, h_out, w_out),
                          mode='trilinear',
                          align_corners=False)[0, 0]
    image = F.interpolate(image[None, None],
                          size=(target_d, target_hw, target_hw),
                          mode='trilinear',
                          align_corners=False)[0, 0]
    return image

Workflow

1. Read PixelSpacing (dy, dx) and SliceThickness (or SpacingBetweenSlices) → dz from the DICOM headers
2. Compute the anisotropy-corrected slice count d_iso = int(dz / dy * d * factor)
3. First interpolate to an isotropic intermediate, then to the fixed network input shape
4. Use trilinear (not nearest, not linear-per-axis) for both passes
5. Cache the per-series spacing alongside the cropped volume to avoid re-reading DICOM headers at training time

Key Decisions

• Physical spacing first, voxel resize second: skipping the spacing correction trains the model on geometrically distorted anatomy.
• `* 0.5` factor: pure dz/dy over-corrects in practice — a damping factor between 0.4 and 0.6 is empirically best for abdominal CT.
• Two-pass interpolation: collapsing the two interpolate calls into one introduces aliasing for high anisotropy ratios.
• `align_corners=False`: the PyTorch default for new code; matches OpenCV/numpy convention.
• Handles missing `dz`: fall back to SpacingBetweenSlices if SliceThickness is absent; some scanners populate only one.

References

• LB 0.55 — 2.5D + 3D sample model